Next Article in Journal
Fabrication of Spiral Low-Cost Microchannel with Trapezoidal Cross Section for Cell Separation Using a Grayscale Approach
Previous Article in Journal
Design and Modeling of a Miniature Hydraulic Motor for Powering a Cutting Tool for Minimally Invasive Procedures
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Micromachines
Volume 14
Issue 7
10.3390/mi14071339
micromachines-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Deep Learning for Microfluidic-Assisted Caenorhabditis elegans Multi-Parameter Identification Using YOLOv7

by
Jie Zhang
1,2,†,
Shuhe Liu
1,†,
Hang Yuan
1,†,
Ruiqi Yong
1,
Sixuan Duan
1,2,
Yifan Li
1,2,
Joseph Spencer
2,
Eng Gee Lim
1,2,*,
Limin Yu
1,2,* and
Pengfei Song
1,2,*
1
School of Advanced Technology, Xi’an Jiaotong-Liverpool University, Suzhou 215123, China
2
Department of Electrical and Electronic Engineering, University of Liverpool, Liverpool L693BX, UK
*
Authors to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Micromachines 2023, 14(7), 1339; https://doi.org/10.3390/mi14071339
Submission received: 12 April 2023 / Revised: 14 June 2023 / Accepted: 27 June 2023 / Published: 29 June 2023
(This article belongs to the Special Issue Celebration of the 3rd Anniversary of the School of Advanced Technology of Xi’an Jiaotong-Liverpool University: Advances in AI and Microengineering)
Browse Figures
Versions Notes

Abstract

:
The Caenorhabditis elegans (C. elegans) is an ideal model organism for studying human diseases and genetics due to its transparency and suitability for optical imaging. However, manually sorting a large population of C. elegans for experiments is tedious and inefficient. The microfluidic-assisted C. elegans sorting chip is considered a promising platform to address this issue due to its automation and ease of operation. Nevertheless, automated C. elegans sorting with multiple parameters requires efficient identification technology due to the different research demands for worm phenotypes. To improve the efficiency and accuracy of multi-parameter sorting, we developed a deep learning model using You Only Look Once (YOLO)v7 to detect and recognize C. elegans automatically. We used a dataset of 3931 annotated worms in microfluidic chips from various studies. Our model showed higher precision in automated C. elegans identification than YOLOv5 and Faster R-CNN, achieving a mean average precision (mAP) at a 0.5 intersection over a union ([email protected]) threshold of 99.56%. Additionally, our model demonstrated good generalization ability, achieving an [email protected] of 94.21% on an external validation set. Our model can efficiently and accurately identify and calculate multiple phenotypes of worms, including size, movement speed, and fluorescence. The multi-parameter identification model can improve sorting efficiency and potentially promote the development of automated and integrated microfluidic platforms.

1. Introduction

Since the 1970s, when Caenorhabditis elegans (C. elegans) was discovered by Sydney Brenner, it has been recognized as an excellent model organism for studying the genetic regulation of organ development and programmed cell death [1,2,3]. C. elegans plays a valuable role in studying genetics, drug development, and cell biology due to its short developmental cycle (3 to 4 days), small size (1 to 1.3 mm), ease of cultivation, and high homology (~65%) of its genetic pathways with humans [4,5,6]. The phenotypes of C. elegans differ (e.g., size and movement speed) at different developmental stages, meeting various specific research demands [7,8].
The size, motility, and fluorescence expression of C. elegans are promising research subjects in specific fields [9]. The varied sizes of C. elegans can facilitate an understanding of complex biological processes, including embryogenesis, development, disease, and aging [10,11]. The motility of worms is closely linked to the genetic mechanisms for maintaining balance and movement, such as muscle contraction and synaptic plasticity [12,13,14]. Otherwise, the fluorescence expression of worms is a useful tool for labeling and observing. For example, green fluorescent protein (GFP) is widely used for optical in vivo imaging and phenotypic observation [11,15].
Efficient and accurate C. elegans sorting has always been a significant challenge in related studies [16]. C. elegans can be sorted based on different phenotypes, such as fluorescence expression, morphology, and motility characteristics [17]. The traditional C. elegans sorting method requires manual selection under a microscope, which is labor-intensive and requires skilled operators [18,19,20]. Later, the complex object parametric analyzer and sorter (COPAS) was developed for efficient and high-throughput worm sorting, utilizing fluorescent signals and other optical features [21,22]. However, the widespread use of COPAS is limited due to its bulkiness, high cost, and operational complexity [23]. In recent years, the advancements of lab-on-a-chip (microfluidics) have made it a promising platform for facilitating various studies, such as microinjection and drug screening [24]. Microfluidic chips are well suited for C. elegans studies due to their highly matched dimensions (in the hundreds of microns to a few millimeters) [25,26,27,28]. In addition, they have the advantages of low cost, good biocompatibility, and easy operation [29,30]. Thanks to software-controlled microscopy and various programmable devices such as micropumps and microvalves, microfluidic chips can rapidly and accurately sort worms [28]. However, these chips still have limitations in integrated, high-content, and multi-parameter sorting [31]. Achieving multi-parameter sorting integrated on a single microfluidic chip is challenging.
Computer vision can automate experimental processes to improve efficiency, enabling a fast response time and customization for multiple complex backgrounds [32,33,34,35,36]. Multiple computer vision techniques have shown good potential for the automated detection of C. elegans [37,38,39,40,41]. Based on deep learning algorithms, a real-time detector system has previously been developed for accurate worm localization, classification, and profile prediction, including the Mask region-based convolutional neural network (R-CNN), Faster R-CNN, and You Only Look Once (YOLO)v5 [41]. The latest version of YOLOv7 has been recognized as one of the most accurate and fastest real-time object detectors, making it ideal for optimizing applications in C. elegans. [42,43,44]. However, applications combining YOLOv7 with microfluidics are yet to be developed.
In this study, we reported a deep learning model using YOLOv7 for microfluidic-assisted C. elegans multi-parameter identification (Figure 1). We extracted C. elegans images in microfluidic chips from various studies as our datasets. Then, we annotated 3931 C. elegans to provide the data required for YOLOv7 training. These datasets, with a large number of annotated images, ensure the effectiveness of our deep learning model. Otherwise, we compared the performance of YOLOv7 to YOLOv5 and Faster R-CNN in detecting C. elegans in microfluidic chips. Furthermore, we examined the generalizability of YOLOv7 using an external validation set. The size and movement speed of C. elegans can be computationally identified by YOLOv7, enabling the classification of multiple phenotypes (size, motility, and fluorescence). Our model achieved a high precision with a mean average precision (mAP) at a 0.5 intersection over a union ([email protected]) threshold of 99.56%. Our model also exhibited good generalization abilities when tested on an external validation set, achieving a [email protected] of 94.21%. Thanks to the identification of multiple parameters with our model, the efficiency of microfluidic-assisted C. elegans research has been improved. Through this functional integration developed by our study, the automation and integration of microfluidic systems are expected to be further improved.

2. Methods

This section discusses the methods used in the following sections: Section 2.1, application scenario; Section 2.2, image acquisition; Section 2.3, dataset preparation; Section 2.4, model configuration and installation; Section 2.5, evaluation metrics; Section 2.6, training configuration; Section 2.7, comparison of YOLOv7, YOLOv5, and Faster R-CNN parameters; and Section 2.8, calculation of worm size and movement speed.

2.1. Application Scenario

Microfluidic chips can be combined well with optical microscopes (e.g., inverted microscopes and fluorescence microscopes) in biology labs for C. elegans research. The combination can potentially improve the efficiency and precision of biological research but requires the assistance of computer algorithms. Thus, our model focuses on the application scenario of microfluidic chips. To capture images, perform morphological and physiological measurements, and sort individual C. elegans, we can combine the model with our previously developed microfluidic chip, which is capable of sequentially loading C. elegans one-by-one (success rate~90.3%) [28]. Our model can detect and identify images of worms moving in microfluidic devices captured by high-speed cameras of microscopes. With image feedback, the model can accurately calculate and determine the phenotype information of each C. elegans (e.g., size, movement speed, and fluorescence) and actively sort the required C. elegans from a mixture of C. elegans with different phenotypes. Our real-time vision-based multi-parameter identification model can improve the automation level of microfluidic systems.

2.2. Image Acquisition

The dataset was obtained from published studies that utilized microfluidic chips for C. elegans sorting. The videos from published studies were downloaded as raw data [45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64]. To improve the model’s robustness and the data’s diversity, we created datasets using reference videos with different imaging conditions from various papers. The acquired videos were sampled at 10-frame intervals to extract cropped worm images and annotate them, resulting in 3931 annotation boxes. After the training phase, the acquired data were sorted into training, testing, and validation sets based on the GFP and WT classification.

2.3. Dataset Preparation

A large amount of annotated data are needed when training the deep learning model. In this study, the worms in each image were annotated manually using the image annotation software LabelImg, which supports the Visual Object Classes (VOC) format [65,66]. This software can generate Extensible Markup Language (XML) files for the model.
To improve the performance of the model training process, we selected C. elegans, which was easily recognizable in the channels as the primary training object. To promote the convergence performance of the model, we carefully selected several worms that were not in the channel as part of the dataset.
Each worm is located within an annotation box, and the number of these boxes determines the dataset size in Figure 2. There were 2426 annotation boxes for the WT category and 1505 for the GFP category. The annotation boxes were categorized into four sets, namely training, testing, validation, and external validation sets. The training and testing sets were used for model training and internal evaluation, while the validation set was utilized for model adjustment. To verify the generalization ability of the model, we also selected 318 annotated worms from separated video sources as the external validation set. The split data ratio was set to 9:1 for training and validation sets to the testing set. The training set to the validation set ratio was also set to 9:1. According to the above ratio, the annotated GFP and WT numbers in each set are shown in Figure 2.

2.4. Model Configuration and Installation

Our model had an overall structure, as shown in Figure 3. The spatial pyramid pooling convolutional set pooling convolutional (SPPCSPC) module is used in object detection algorithms, specifically in the YOLOv7 model. The SPPCSPC module refers to a Cross Stage Partial Network (CSPNet) with a spatial pyramid pooling (SPP) block. The SPP block pools the input tensor with various kernel sizes and concatenates the results. It shows the difference in maximum pooling. Different pooling amounts correspond to different targets. Moreover, the SPP block has four different scales of maximum pooling (the resolutions of 1 × 1, 5 × 5, 9 × 9, and 13 × 13) with four kinds of perceptual fields, distinguishing large and small targets.
The innovative transition block (TB) is used for down-sampling. In convolutional neural networks, two commonly used transition modules for down-sampling are a convolutional layer with a kernel size of 3 × 3 and a stride of 2 × 2, or a max pooling layer with a stride of 2 × 2. YOLOv7 combines these two transition modules into a TB with two branches. The left branch of the TB is a 2 × 2 stride max pooling followed by a 1 × 1 convolutional layer. The right branch of the TB is a 1 × 1 convolutional layer followed by a convolutional layer with a kernel size of 3 × 3 and a stride of 2 × 2. The results from both branches are stacked together when outputting.
The overall structure of our model utilized multi-branch stacking modules, represented by the MCB in Figure 3 and illustrated in Figure 4. In this module, the input of the final stacking module contained multiple branches. From left to right, the four branches shown in Figure 4 consist of 1, 1, 3, and 5 CBS, respectively. After these branches were stacked, they conducted another CBS for feature integration. The dense stacking corresponded to a dense residual structure that included internal residual blocks. While increasing the depth of the neural network can optimize and improve model accuracy, it may lead to vanishing gradient problems. Internal residual blocks with skip connections can mitigate this issue.
YOLOv7 constructs a feature pyramid network (FPN) to enhance the feature extraction. The FPN is located between convolution (Conv) and reparameterized convolution (RepConv) structures in Figure 3, including Conv but excluding RepConv. In the feature utilization stage, YOLOv7 extracts three feature layers at different positions in the backbone for object detection: the middle, lower middle, and bottom layers [67]. Each feature point on each feature layer in YOLOv7 has three prior boxes. The channel number of each feature layer can be adjusted using convolution so that the final number of channels is proportional to the number of distinct categories. These three layers have different shapes when the input is (640, 640, 3): feat = (80, 80, 512), feat2 = (40, 40, 1024), and feat3 = (20, 20, 1024). The FPN is constructed using these three feature layers, combing feature layers of various shapes to obtain better features. In YOLOv7, the FPN produces three enhanced features with shapes of (20, 20, 512), (40, 40, 256), and (80, 80, 128), which are passed to the YOLO detection head for result prediction. Unlike previous YOLO models, YOLOv7 employs a RepConv structure before the YOLO detection head. RepConv structure introduces new residual structures uniquely designed to aid in training. During prediction, these structures can be effectively equivalent to a standard 3 × 3 convolution. Furthermore, they can reduce network complexity while maintaining prediction performance.

2.5. Evaluation Metrics

The performance of YOLOv7 was evaluated using several metrics. Recall is the ratio of all true positives (objects correctly identified) to the total of true positives and false negatives (objects missed). Precision is the ratio of true positives to the total of true and false positives (objects misidentified). AP is the area under the precision–recall curve, and its average value for each class is the mAP. The intersection over union (IoU) threshold was used to measure the degree of overlap between the predicted bounding box and the actual ground truth box. It is defined mathematically as the ratio of the intersection area of the two boxes to their union. The F1 score, the harmonic mean of precision, and recall were used to evaluate the overall performance of the model. Moreover, the detection speed of the model in frames per second (FPS) on the tracked video was evaluated. Finally, the IoU threshold for the tracked frame was set to 0.5.
The training process for this model involved two stages: the frozen stage and the unfrozen stage. During the frozen stage, the backbone was frozen and the feature extraction network remained unchanged. The training was conducted for 50 epochs. The backbone was unfrozen and the feature extraction network was modified during the unfrozen stage. The training was conducted for 300 epochs. The learning rate of the model was dynamically adjusted, with a range of 1 × 10−4 to 1 × 10−2. The stochastic gradient descent optimizer was used with a weight decay of 5 × 10−4. Real-time training progress monitoring was achieved by performing an average evaluation every 10 epochs, which provided valuable information on the model performance. This information helped to identify any overfitting or underfitting issues and enabled the adjustment of hyperparameters or the training process to improve model performance. For example, if the precision of the model decreases, reducing the learning rate or increasing the number of epochs may help. Alternatively, if the model is overfitting, regularization techniques may be applied to prevent the memorization of training data. By conducting regular evaluations, the training process can be optimized for maximum model performance.
The YOLO detection head generates prediction results of three feature layers with shapes of (N, 20, 20, 255), (N, 40, 40, 255), and (N, 80, 80, 255), respectively. However, these predictions do not correspond to the positions of the final predicted boxes on the image. They need to be decoded to complete the process. In YOLOv7, each feature point on each feature layer has three anchor boxes. The final 255 values of each feature layer can be split into three sets of 85 parameters, which correspond to the 3 anchor boxes. These 85 parameters can be reshaped into (N, 20, 20, 3, 85), (N, 40, 40, 3, 85), and (N, 80, 80, 3, 85) for the three feature layers, respectively. Among these parameters, 85 can be further divided into 4 + 1 + 80. The first 4 parameters are used to adjust the regression parameters of each feature point and obtain the predicted box. The 5th parameter determines whether each feature point contains an object. The last 80 parameters identify the object type in each feature point.
First, the feature layer divides the image into grid cells in the object detection model. Then, any grid cell that falls within an object’s corresponding box is used to predict the object. Next, score filtering and non-maximum suppression (NMS) filtering are performed to reduce the number of prediction boxes. Score filtering selects boxes with scores that meet the confidence threshold, thus significantly reducing the number of boxes to be processed during the overlap removal. Meanwhile, NMS selects the box with the highest score for a certain category within a certain region. To apply NMS to each category independently, each category is looped through, and the boxes with lower scores for the same category are filtered out. After filtering, the boxes for each category are sorted in descending order of their scores. Then, the box with the highest score is removed each time, and its overlap with all other predicted boxes is calculated. Boxes with high overlap are removed. Finally, the remaining prediction boxes are used for the final predictions. In our case, we used a score filtering value of 0.5 and an NMS parameter (nms_iou) of 0.3.

2.6. Training Configuration

We trained our model on the Windows 10 operating system using an NVIDIA GeForce RTX 3090 GPU with 32 GB RAM. We implemented our model using version 1.7.1 of the PyTorch machine learning framework. Moreover, our model depended on NVIDIA’s CUDA (version 11.1) and CUDNN libraries for support.

2.7. Comparison of the YOLOv7, YOLOv5, and Faster R-CNN Parameters

Our study employed YOLOv7 to identify the WT and GFP datasets and compare its performance against YOLOv5 and Faster R-CNN. We have used the best hyperparameters for three models. The input size of the images for both YOLOv7 and YOLOv5 was 640 × 640, while the size for Faster R-CNN was 600 × 600. Additionally, the batch size for training three models was the same and equaled 3. Table 1 summarizes the parameters used to train the three models.

2.8. Calculation of Worm Size and Movement Speed

To obtain the size of the C. elegans, we determined the actual size of each pixel point based on the scale, and then computed the pixel positions of the worm’s boundaries. The number of occupied pixels was used to determine the worm size. To determine the movement speed of the C. elegans, we recorded the Greenwich time and left boundary of the frame for each detected image, and calculated the difference in time and boundary position between adjacent frames. We conducted tests using two distinct microfluidic chips selected from the dataset videos. Then, we converted pixel numbers into actual lengths of worms based on microfluidic chip lengths of 4.6875 × 103 μm (chip 1) and 3.4286 × 103 μm (chip 2), and resolutions of 863 × 725 pixels (chip 1) and 318 × 232 pixels (chip 2). Consequently, we obtained the worm size and movement speed.

3. Results

We performed the process shown in Figure 5 using YOLOv7 to detect C. elegans during sorting in microfluidic chips. We compared the performance of YOLOv7, YOLOv5, and Faster R-CNN in Table 2. YOLOv7 achieved the highest mAP of 99.56% for detecting the WT and GFP datasets, outperforming YOLOv5 and Faster R-CNN.
We compared the performance of YOLOv7, YOLOv5, and Faster R-CNN in C. elegans detection, and the results of the evaluation metrics are shown in Table 3. The specific results of the testing set are presented in Figures S1–S9 of the Supplementary Materials. The three models used the same number of images for object detection, with 2426 and 1505 images for WT and GFP, respectively. YOLOv7 showed the fastest detection speed among the three models, with an FPS of 52.112. Moreover, YOLOv7 exhibited higher precision, achieving 100% and 97.39% for detecting WT and GFP, respectively. Although Faster R-CNN showed a slightly better recall of 99.35% in GFP detection than the other two models, overall, YOLOv7 demonstrated higher F1 scores in both WT and GFP detection.
An external validation set can help ensure our model can generalize to new data rather than overfitting the training data. Therefore, we compared the performance of YOLOv7, YOLOv5, and Faster R-CNN on an external validation set to evaluate their generalization ability. The results of the evaluation metrics are shown in Table 4. The specific results of the external validation set are presented in Figures S10–S18 of the Supplementary Materials. The three models used the same number of annotation boxes from the external validation set, with 154 and 164 annotated WT and GFP C. elegans, respectively. YOLOv7 showed the fastest detection speed among the three models, with an FPS of 52.508. In addition, YOLOv7 exhibited the highest precision for both WT and GFP, reaching 95.24% and 89.47%, respectively. YOLOv7 also showed the highest mAP of 94.21%, indicating that our model has robust detection and classification capabilities for C. elegans. Otherwise, YOLOv7 had the highest recall of 89.47% for GFP, although its recall for WT was slightly lower than that of Faster R-CNN but still better than that of YOLOv5. Overall, YOLOv7 has a better balance between precision and recall, with F1 scores of 0.95 and 0.89 for WT and GFP, respectively.
We applied our model to recognize C. elegans automatically in microfluidic chips from videos. The mixed populations of WT and GFP C. elegans were imaged in the microfluidic chips under an optical microscope. Our model could identify the fluorescent phenotype of these worms (WT or GFP) and calculate their size and speed (Figure 6). In practical application scenarios, microfluidic chips can be combined with our model to sort GFP and WT worms into different channels to obtain a single target population. Moreover, combined with different microfluidic chip designs, the size and speed output results can also serve as sorting criteria. As shown in Table 5, we employed twenty-eight worms for identification in our model, including twenty-four WT and four GFP worms from chip 1 (Worm1–11 and Worm25–28) and chip 2 (Worm12–24). The largest sized worm was 94,394.17 μm2 and the smallest sized was 11,547.36 μm2. The fastest worm had a speed of 2827.30 μm/s, while the slowest was 46.13 μm/s. Since the recognition was performed on videos, the speed 0 is invalid due to interference from the video frame rate. Variations in C. elegans’ movement speed may arise due to differences in video frame rates, air pump settings, and other experimental conditions of the microfluidic chips [68]. However, it is important to note that these factors do not invalidate the obtained results. On the contrary, they showcase the adaptability of our model to diverse microfluidic application environments, highlighting the robustness and effectiveness of our approach.

4. Discussion

Researchers are increasingly using object detectors to study C. elegans. While fluorescent techniques are helpful for tracking and sorting C. elegans, sorting typically involves multiple parameters. Therefore, combining deep learning models with microfluidic chips is a critical way to achieve platform automation and improve the efficiency of multi-parameter identification. We developed a target detection model to identify WT and GFP C. elegans in microfluidic chips and generate their size and speed information. This approach can improve the precision and efficiency of microfluidic chips during C. elegans sorting, reducing errors associated with manual counting.
We extracted C. elegans in microfluidic chips from videos and annotated them to create a dataset. We trained three commonly used detectors (YOLOv7, YOLOv5, and Faster R-CNN) using the dataset and validated their performance using an external validation set. The comparison results of their performance showed that YOLOv7 outperformed the others in terms of recall, precision, and generalization ability. Therefore, we ultimately chose to apply YOLOv7 in our study. Our study provided a dataset of thousands of annotated worms. Additionally, we developed an engineering prototype that combines microfluidic chips and deep learning technology for efficient C. elegans sorting. The model provided a solution to identify multiple parameters of worms in microfluidic chips.
In microfluidic chips, worms are introduced into microchannels, where they experience fluid flow and confinement. These microfluidic channels are typically designed with specific geometries and dimensions to create controlled flow conditions [17]. The worms are not completely immobilized but rather constrained to move and navigate within the available space, interacting with the fluid and channel walls. Some microfluidic chips are also equipped with active forces, such as different types of pumps. The assistance of these active forces helps ensure efficient sorting and prevent any potential interference or blockage within the chips. The applied active forces and specific channel geometries influence worm behaviors, allowing them to be sorted into different bifurcated channels based on predefined criteria such as size and speed. The movement of worms may exhibit inherent randomness and a tendency to interact with the channel sidewalls, displaying exploratory behavior [68]. In further research, integrating our algorithm and microfluidic chips is expected to measure, control, and predict worm behaviors precisely within microchannels, contributing to a more reliable analysis.
We can also easily implement our identification model for efficient and automated multi-parameter C. elegans sorting with our developed C. elegans soring microfluidic systems, which are capable of continuously loading C. elegans with a 90.3% success rate [28]. Based on the current performance of our model, we anticipate achieving a high recognition success rate and target population purity in practical applications. For example, C. elegans is commonly used in drug screening and exhibits a wide range of variations in movement. These variations can arise from various factors such as individual differences, drug effects, and environmental conditions [69]. Addressing the variations in worm movement is crucial for accurately assessing the effects of drugs in experiments. By leveraging our model, the efficient and automated analysis of worm locomotion under different drug treatments can be performed, thereby accelerating the drug screening process and enhancing screening accuracy. Furthermore, precise measurements of worm movement variations can be used to evaluate the toxicological effects of drugs and provide robust support for drug safety assessments. We will further explore the potential applications of our deep learning model in these areas.

5. Conclusions

C. elegans is a model organism with great potential in human disease and genetics research. The efficient and accurate sorting of C. elegans has always been a challenge in related research. Microfluidic chips are promising for C. elegans sorting due to their low cost, good biocompatibility, and simple operation, compared with the cumbersome and inefficient manual operation and expensive COPAS. However, automated C. elegans sorting with multiple parameters requires efficient identification technology due to the different research demands for worm phenotypes. Therefore, we extracted and annotated 3931 worms from videos as a dataset and selected YOLOv7 through the training, comparison, and external validation of three detectors (YOLOv7, YOLOv5, and Faster R-CNN). YOLOv7 showed high precision in worm recognition, achieving an [email protected] of 99.56%. In addition, our model has a good generalization ability, achieving an [email protected] of 94.21% on the external validation set while efficiently and accurately identifying multiple C. elegans phenotypes, including size, movement speed, and fluorescence. This study comprehensively evaluated the performance of YOLOv7 as a C. elegans recognition detector. Otherwise, we established a new engineering prototype that combines microfluidics and deep learning technology. We will further develop highly integrated and automated microfluidic chips based on our model to better serve C. elegans-related research.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/mi14071339/s1, Specific results of the testing set. Figure S1. (a) The mAP, (b) the log-average miss rate, and (c) the ground-truth results in Faster R-CNN. Figure S2. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of WT C. elegans in Faster R-CNN. Figure S3. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of GFP C. elegans in Faster R-CNN. Figure S4. (a) The mAP, (b) the log-average miss rate, and (c) the ground-truth results in YOLOv5. Figure S5. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of WT C. elegans in YOLOv5. Figure S6. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of GFP C. elegans in YOLOv5. Figure S7. (a) The mAP, (b) the log-average miss rate, and (c) the ground-truth results in YOLOv7. Figure S8. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of WT C. elegans in YOLOv7. Figure S9. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of GFP C. elegans in YOLOv7. Specific results of the external validation set. Figure S10. (a) The mAP, (b) the log-average miss rate, and (c) the ground-truth results in Faster R-CNN. Figure S11. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of WT C. elegans in Faster R-CNN. Figure S12. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of GFP C. elegans in Faster R-CNN. Figure S13. (a) The mAP, (b) the log-average miss rate, and (c) the ground-truth results in YOLOv5. Figure S14. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of WT C. elegans in YOLOv5. Figure S15. (a) The recall, (b) the F1 score, (c) precision, and (d) the AP results of GFP C. elegans in YOLOv5. Figure S16. (a) The mAP, (b) the log-average miss rate, and (c) the ground-truth results in YOLOv7. Figure S17. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of WT C. elegans in YOLOv7. Figure S18. (a) The recall, (b) the F1 score, (c) the precision, and (d) the AP results of GFP C. elegans in YOLOv7.

Author Contributions

Conceptualization, J.Z., S.L. and P.S.; Data curation, J.Z. and S.L.; Formal analysis, J.Z., S.L., H.Y. and R.Y.; Methodology, J.Z. and S.L.; Resources, J.Z., S.L. and H.Y.; Supervision, J.S., E.G.L., L.Y. and P.S.; Writing—original draft, J.Z., S.L., H.Y. and P.S.; Writing—review and editing, J.Z., S.L., H.Y., R.Y., S.D., Y.L., J.S., E.G.L., L.Y. and P.S. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Jiangsu Science and Technology Programme – Young Scholar (BK20200251), XJTLU Research Development Fund (RDF-18-02-20), and Xi’an Jiaotong–Liverpool University to Sixuan Duan (PGRS1912019). This work is also partially supported by the XJTLU AI University Research Centre and Jiangsu Province Engineering Research Centre of Data Science and Cognitive Computation at XJTLU.

Data Availability Statement

The code used and the trained models in this study are freely available and are hosted in a dedicated open-source repository on GitHub [70].

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Brenner, S. The genetics of Caenorhabditis elegans. Genetics 1974, 77, 71–94. [Google Scholar] [CrossRef] [PubMed]
  2. Markaki, M.; Tavernarakis, N. Modeling human diseases in Caenorhabditis elegans. Biotechnol. J. 2010, 5, 1261–1276. [Google Scholar] [CrossRef] [PubMed]
  3. Corsi, A.K. A biochemist’s guide to Caenorhabditis elegans. Anal. Biochem. 2006, 359, 1–17. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  4. Midkiff, D.; San-Miguel, A. Microfluidic Technologies for High Throughput Screening through Sorting and On-Chip Culture of C. elegans. Molecules 2019, 24, 4292. [Google Scholar] [CrossRef] [Green Version]
  5. Braungart, E.; Gerlach, M.; Riederer, P.; Baumeister, R.; Hoener, M.C. Caenorhabditis elegans MPP+ model of Parkinson’s disease for high-throughput drug screenings. Neurodegener. Dis. 2004, 1, 175–183. [Google Scholar] [CrossRef]
  6. Youssef, K.; Tandon, A.; Rezai, P. Studying Parkinson’s disease using Caenorhabditis elegans models in microfluidic devices. Integr. Biol. 2019, 11, 186–207. [Google Scholar] [CrossRef]
  7. Sinha, D.B.; Pincus, Z.S. High temporal resolution measurements of movement reveal novel early-life physiological decline in C. elegans. PLoS ONE 2022, 17, e0257591. [Google Scholar] [CrossRef]
  8. Atakan, H.B.; Ayhan, F.; Gijs, M.A.M. PDMS filter structures for size-dependent larval sorting and on-chip egg extraction of C. elegans. Lab Chip 2020, 20, 155–167. [Google Scholar] [CrossRef]
  9. Zhao, Z.K.; Liu, J.; Wang, J.J.; Chen, G.; Li, C.Q.; Zhao, Z.; Jiang, M. Age estimation of Caenorhabditis elegans with label distribution learning. In Proceedings of the 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Man, Prague, Czech Republic, 9–12 October 2022; pp. 3137–3142. [Google Scholar]
  10. Girard, L.; Fiedler, T.; Harris, T.; Carvalho, F.; Antoshechkin, I.; Han, M.; Sternberg, P.; Stein, L.; Chalfie, M. WormBook: The online review of Caenorhabditis elegans biology. Nucleic Acids Res. 2007, 35, D472–D475. [Google Scholar] [CrossRef]
  11. Huang, C.; Xiong, C.; Kornfeld, K. Measurements of age-related changes of physiological processes that predict lifespan of Caenorhabditis elegans. Proc. Natl. Acad. Sci. USA 2004, 101, 8084–8089. [Google Scholar] [CrossRef] [Green Version]
  12. Sakamoto, K.; Soh, Z.; Suzuki, M.; Iino, Y.; Tsuji, T. Forward and backward locomotion patterns in C. elegans generated by a connectome-based model simulation. Sci. Rep. 2021, 11, 13737. [Google Scholar] [CrossRef]
  13. Pierce-Shimomura, J.T.; Chen, B.L.; Mun, J.J.; Ho, R.; Sarkis, R.; McIntire, S.L. Genetic analysis of crawling and swimming locomotory patterns in C. elegans. Proc. Natl. Acad. Sci. USA 2008, 105, 20982–20987. [Google Scholar] [CrossRef] [Green Version]
  14. Maulik, M.; Mitra, S.; Bult-Ito, A.; Taylor, B.E.; Vayndorf, E.M. Behavioral Phenotyping and Pathological Indicators of Parkinson’s Disease in C. elegans Models. Front. Genet. 2017, 8, 77. [Google Scholar] [CrossRef] [Green Version]
  15. Rapti, G. A perspective on C. elegans neurodevelopment: From early visionaries to a booming neuroscience research. J. Neurogenet. 2020, 34, 259–272. [Google Scholar] [CrossRef]
  16. Koopman, M.; Peter, Q.; Seinstra, R.I.; Perni, M.; Vendruscolo, M.; Dobson, C.M.; Knowles, T.P.J.; Nollen, E.A.A. Assessing motor-related phenotypes of Caenorhabditis elegans with the wide field-of-view nematode tracking platform. Nat. Protoc. 2020, 15, 2071–2106. [Google Scholar] [CrossRef]
  17. Yuan, H.; Yuan, W.; Duan, S.; Jiao, K.; Zhang, Q.; Lim, E.; Chen, M.; Zhao, C.; Pan, P.; Liu, X.; et al. Microfluidic Assisted Caenorhabditis elegans Sorting: Current Status and Future Prospects. Cyborg Bionic Syst. 2023, 4, 0011. [Google Scholar] [CrossRef]
  18. Han, B.; Kim, D.; Hyun Ko, U.; Shin, J.H. A sorting strategy for C. elegans based on size-dependent motility and electrotaxis in a micro-structured channel. Lab Chip 2012, 12, 4128–4134. [Google Scholar] [CrossRef]
  19. Chung, K.; Crane, M.M.; Lu, H. Automated on-chip rapid microscopy, phenotyping and sorting of C. elegans. Nat. Methods 2008, 5, 637–643. [Google Scholar] [CrossRef]
  20. Stiernagle, T. Maintenance of C. elegans. WormBook: Online Review of C. elegans Biology; WormBook: Pasadena, CA, USA, 2006; Volume 11, pp. 1–11. [Google Scholar] [CrossRef] [Green Version]
  21. Pulak, R. Techniques for analysis, sorting, and dispensing of C. elegans on the COPAS flow-sorting system. Methods Mol. Biol. 2006, 351, 275–286. [Google Scholar] [CrossRef]
  22. Doitsidou, M.; Flames, N.; Lee, A.C.; Boyanov, A.; Hobert, O. Automated screening for mutants affecting dopaminergic-neuron specification in C. elegans. Nat. Methods 2008, 5, 869–872. [Google Scholar] [CrossRef] [Green Version]
  23. Clark, A.S.; Huayta, J.; Arulalan, K.S.; San-Miguel, A. Chapter 13—Microfluidic devices for imaging and manipulation of C. elegans. In Micro and Nano Systems for Biophysical Studies of Cells and Small Organisms; Liu, X., Sun, Y., Eds.; Academic Press: Cambridge, MA, USA, 2021; pp. 295–321. [Google Scholar] [CrossRef]
  24. Levine, E.; Lee, K.S. Microfluidic approaches for Caenorhabditis elegans research. Anim. Cells Syst. 2020, 24, 311–320. [Google Scholar] [CrossRef] [PubMed]
  25. Pan, P.; Qin, Z.; Sun, W.; Zhou, Y.; Wang, S.; Song, P.; Wang, Y.; Ru, C.; Wang, X.; Calarco, J.; et al. A spiral microfluidic device for rapid sorting, trapping, and long-term live imaging of Caenorhabditis elegans embryos. Microsyst. Nanoeng. 2023, 9, 17. [Google Scholar] [CrossRef] [PubMed]
  26. Xianke, D.; Pengfei, S.; Liu, X. An automated robotic system for high-speed microinjection of Caenorhabditis elegans. In Proceedings of the 2015 IEEE International Conference on Robotics and Automation (ICRA), Seattle, WA, USA, 26–30 May 2015; pp. 996–1001. [Google Scholar]
  27. Dong, X.; Song, P.; Liu, X. Automated Robotic Microinjection of the Nematode Worm Caenorhabditis elegans. IEEE Trans. Autom. Sci. Eng. 2021, 18, 850–859. [Google Scholar] [CrossRef]
  28. Dong, X.; Song, P.; Liu, X. An Automated Microfluidic System for Morphological Measurement and Size-Based Sorting of C. elegans. IEEE Trans. NanoBiosci. 2019, 18, 373–380. [Google Scholar] [CrossRef]
  29. Song, P.; Dong, X.; Liu, X. A microfluidic device for automated, high-speed microinjection of Caenorhabditis elegans. Biomicrofluidics 2016, 10, 011912. [Google Scholar] [CrossRef] [Green Version]
  30. Song, P.; Zhang, W.; Sobolevski, A.; Bernard, K.; Hekimi, S.; Liu, X. A microfluidic device for efficient chemical testing using Caenorhabditis elegans. Biomed. Microdevices 2015, 17, 38. [Google Scholar] [CrossRef]
  31. Sivagnanam, V.; Gijs, M.A. Exploring living multicellular organisms, organs, and tissues using microfluidic systems. Chem. Rev. 2013, 113, 3214–3247. [Google Scholar] [CrossRef]
  32. Rico-Guardiola, E.J.; Layana-Castro, P.E.; García-Garví, A.; Sánchez-Salmerón, A.-J. Caenorhabditis elegans Detection Using YOLOv5 and Faster R-CNN Networks. In Proceedings of the Optimization, Learning Algorithms and Applications, Póvoa de Varzim, Portugal, 24–25 October 2022; pp. 776–787. [Google Scholar]
  33. Yuan, H.; Zhang, W. A Novel Hedgehog-Inspired Pin-Array Robot Hand with Multiple Magnetic Pins for Adaptive Grasping. In Proceedings of the Intelligent Robotics and Applications: 12th International Conference (ICIRA 2019), Shenyang, China, 8–11 August 2019; Part V 12. pp. 684–695. [Google Scholar]
  34. Gallo, I.; Rehman, A.U.; Dehkordi, R.H.; Landro, N.; La Grassa, R.; Boschetti, M. Deep Object Detection of Crop Weeds: Performance of YOLOv7 on a Real Case Dataset from UAV Images. Remote Sens. 2023, 15, 539. [Google Scholar] [CrossRef]
  35. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Region-Based Convolutional Networks for Accurate Object Detection and Segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 2016, 38, 142–158. [Google Scholar] [CrossRef]
  36. Duan, S.; Cai, T.; Zhu, J.; Yang, X.; Lim, E.G.; Huang, K.; Hoettges, K.; Zhang, Q.; Fu, H.; Guo, Q.; et al. Deep learning-assisted ultra-accurate smartphone testing of paper-based colorimetric ELISA assays. Anal. Chim. Acta 2023, 1248, 340868. [Google Scholar] [CrossRef]
  37. Rodriguez, M.; Snoek, L.B.; De Bono, M.; Kammenga, J.E. Worms under stress: C. elegans stress response and its relevance to complex human disease and aging. Trends Genet. 2013, 29, 367–374. [Google Scholar] [CrossRef]
  38. Angstman, N.B.; Frank, H.-G.; Schmitz, C. Advanced Behavioral Analyses Show that the Presence of Food Causes Subtle Changes in C. elegans Movement. Front. Behav. Neurosci. 2016, 10, 60. [Google Scholar] [CrossRef] [Green Version]
  39. Mujika, A.; Leškovský, P.; Álvarez, R.; Otaduy, M.A.; Epelde, G. Modeling Behavioral Experiment Interaction and Environmental Stimuli for a Synthetic C. elegans. Front. Neuroinform. 2017, 11, 71. [Google Scholar] [CrossRef] [Green Version]
  40. Banerjee, S.C.; Khan, K.A.; Sharma, R. Deep-Worm-Tracker: Deep Learning Methods for Accurate Detection and Tracking for Behavioral Studies in C. elegans. bioRxiv 2022. [Google Scholar] [CrossRef]
  41. Fudickar, S.; Nustede, E.J.; Dreyer, E.; Bornhorst, J. Mask R-CNN Based C. elegans Detection with a DIY Microscope. Biosensors 2021, 11, 257. [Google Scholar] [CrossRef]
  42. Wang, C.-Y.; Bochkovskiy, A.; Liao, H.-Y. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 18–22 June 2023; pp. 7464–7475. [Google Scholar]
  43. Ge, Z.; Liu, S.; Li, Z.; Yoshie, O.; Sun, J. OTA: Optimal Transport Assignment for Object Detection. In Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 20–25 June 2021; pp. 303–312. [Google Scholar]
  44. Ge, Z.; Liu, S.; Wang, F.; Li, Z.; Sun, J. YOLOX: Exceeding YOLO Series in 2021. arXiv 2021, arXiv:2107.08430. [Google Scholar]
  45. Zhang, N.; Membreno, E.; Raj, S.; Zhang, H.; Khan, L.A.; Gobel, V. The C. elegans Excretory Canal as a Model for Intracellular Lumen Morphogenesis and In Vivo Polarized Membrane Biogenesis in a Single Cell: Labeling by GFP-fusions, RNAi Interaction Screen and Imaging. J. Vis. Exp. 2017, 128, 56101. [Google Scholar] [CrossRef]
  46. Zhang, N.; Khan, L.A.; Membreno, E.; Jafari, G.; Yan, S.; Zhang, H.; Gobel, V. The C. elegans Intestine As a Model for Intercellular Lumen Morphogenesis and In Vivo Polarized Membrane Biogenesis at the Single-Cell Level: Labeling by Antibody Staining, RNAi Loss-of-Function Analysis and Imaging. J. Vis. Exp. 2017, 128, e56100. [Google Scholar] [CrossRef]
  47. Squiban, B.; Belougne, J.; Ewbank, J.; Zugasti, O. Quantitative and automated high-throughput genome-wide RNAi screens in C. elegans. J. Vis. Exp. 2012, 60, e3448. [Google Scholar] [CrossRef] [Green Version]
  48. Nussbaum-Krammer, C.I.; Neto, M.F.; Brielmann, R.M.; Pedersen, J.S.; Morimoto, R.I. Investigating the spreading and toxicity of prion-like proteins using the metazoan model organism C. elegans. J. Vis. Exp. 2015, 95, e52321. [Google Scholar] [CrossRef]
  49. Yan, Y.; Ng, L.F.; Ng, L.T.; Choi, K.B.; Gruber, J.; Bettiol, A.A.; Thakor, N.V. A continuous-flow C. elegans sorting system with integrated optical fiber detection and laminar flow switching. Lab Chip 2014, 14, 4000–4006. [Google Scholar] [CrossRef] [PubMed]
  50. Berkowitz, L.A.; Hamamichi, S.; Knight, A.L.; Harrington, A.J.; Caldwell, G.A.; Caldwell, K.A. Application of a C. elegans dopamine neuron degeneration assay for the validation of potential Parkinson’s disease genes. J. Vis. Exp. 2008, 17, 835. [Google Scholar] [CrossRef] [Green Version]
  51. Yuan, J.; Zhou, J.; Raizen, D.M.; Bau, H.H. High-throughput, motility-based sorter for microswimmers such as C. elegans. Lab Chip 2015, 15, 2790–2798. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  52. Suzuki, M.; Sakashita, T.; Funayama, T. Immobilization of Live Caenorhabditis elegans Individuals Using an Ultra-Thin Polydimethylsiloxane Microfluidic Chip with Water Retention. J. Vis. Exp. 2019, 145, e59008. [Google Scholar] [CrossRef] [Green Version]
  53. Day, N.J.; Wang, X.; Voronina, E. In Situ Detection of Ribonucleoprotein Complex Assembly in the C. elegans Germline using Proximity Ligation Assay. J. Vis. Exp. 2020, 159, e60982. [Google Scholar] [CrossRef]
  54. Gaffney, C.J.; Bass, J.J.; Barratt, T.F.; Szewczyk, N.J. Methods to assess subcellular compartments of muscle in C. elegans. J. Vis. Exp. 2014, 93, e52043. [Google Scholar] [CrossRef] [Green Version]
  55. Bijwadia, S.R.; Morton, K.; Meyer, J.N. Quantifying Levels of Dopaminergic Neuron Morphological Alteration and Degeneration in Caenorhabditis elegans. J. Vis. Exp. 2021, 177, e62894. [Google Scholar] [CrossRef]
  56. Lazaro-Pena, M.I.; Cornwell, A.B.; Samuelson, A.V. Quantifying Tissue-Specific Proteostatic Decline in Caenorhabditis elegans. J. Vis. Exp. 2021, 175, e61100. [Google Scholar] [CrossRef]
  57. Cinkornpumin, J.K.; Hong, R.L. RNAi mediated gene knockdown and transgenesis by microinjection in the necromenic Nematode Pristionchus pacificus. J. Vis. Exp. 2011, 56, e3270. [Google Scholar] [CrossRef] [Green Version]
  58. Naji, A.; Al Hatem, A.; van der Hoeven, R. Studying Oxidative Stress Caused by the Mitis Group Streptococci in Caenorhabditis elegans. J. Vis. Exp. 2019, 145, e59301. [Google Scholar] [CrossRef]
  59. Zeng, F.; Rohde, C.B.; Yanik, M.F. Sub-cellular precision on-chip small-animal immobilization, multi-photon imaging and femtosecond-laser manipulation. Lab Chip 2008, 8, 653–656. [Google Scholar] [CrossRef]
  60. Garcia, A.M.; Ladage, M.L.; Padilla, P.A. Use of time lapse microscopy to visualize anoxia-induced suspended animation in C. elegans embryos. J. Vis. Exp. 2012, 70, e4319. [Google Scholar] [CrossRef] [Green Version]
  61. Zhu, T.; Zhu, L.; Li, Y.; Chen, X.; He, M.; Li, G.; Zhang, H.; Gao, S.; Fei, P. High-speed large-scale 4D activities mapping of moving C. elegans by deep-learning-enabled light-field microscopy on a chip. Sens. Actuators B Chem. 2021, 348, 130638. [Google Scholar] [CrossRef]
  62. Hu, L.; Ge, A.; Wang, X.; Wang, S.; Yue, X.; Wang, J.; Feng, X.; Du, W.; Liu, B.F. Real-time monitoring of immune responses under pathogen invasion and drug interference by integrated microfluidic device coupled with worm-based biosensor. Biosens. Bioelectron. 2018, 110, 233–238. [Google Scholar] [CrossRef]
  63. Yan, Y.; Boey, D.; Ng, L.T.; Gruber, J.; Bettiol, A.; Thakor, N.V.; Chen, C.H. Continuous-flow C. elegans fluorescence expression analysis with real-time image processing through microfluidics. Biosens. Bioelectron. 2016, 77, 428–434. [Google Scholar] [CrossRef]
  64. Zhang, J.; Hartman, J.H.; Chen, C.; Yang, S.; Li, Q.; Tian, Z.; Huang, P.H.; Wang, L.; Meyer, J.N.; Huang, T.J. Fluorescence-based sorting of Caenorhabditis elegans via acoustofluidics. Lab Chip 2020, 20, 1729–1739. [Google Scholar] [CrossRef]
  65. Alon, H.D.; Ligayo, M.A.D.; Misola, M.A.; Sandoval, A.A.; Fontanilla, M.V. Eye-Zheimer: A Deep Transfer Learning Approach of Dementia Detection and Classification from NeuroImaging. In Proceedings of the 2020 IEEE 7th International Conference on Engineering Technologies and Applied Sciences (ICETAS), Kuala Lumpur, Malaysia, 18–20 December 2020; pp. 1–4. [Google Scholar]
  66. Kaufmane, E.; Sudars, K.; Namatēvs, I.; Kalniņa, I.; Judvaitis, J.; Balašs, R.; Strautiņa, S. QuinceSet: Dataset of annotated Japanese quince images for object detection. Data Brief 2022, 42, 108332. [Google Scholar] [CrossRef]
  67. Ding, X.; Zhang, X.; Ma, N.; Han, J.; Ding, G.; Sun, J. RepVGG: Making VGG-style ConvNets Great Again. In Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 20–25 June 2021; pp. 13733–13742. [Google Scholar]
  68. Santosh, P.; Andrew, J.; Roy, L.; Archana, P. Decision-making by nematodes in complex microfluidic mazes. Adv. Biosci. Biotechnol. 2011, 2, 409–415. [Google Scholar] [CrossRef] [Green Version]
  69. Roy, L.; Archana, P.; Santosh, P. Microfluidics-enabled method to identify modes of Caenorhabditis elegans paralysis in four anthelmintics. Biomicrofluidics 2013, 7, 064103. [Google Scholar]
  70. Zhang, J.; Liu, S.; Yuan, H.; Yong, R.; Duan, S.; Li, Y.; Lim, E.G.; Yu, L.; Spencer, J.; Song, P. Deep Learning for Microfluidic Assisted Caenorhabditis elegans Multi-parameter Identification Using YOLOv7 [Code Used and Trained Models]. GitHub. 2023. Available online: https://github.com/EnderHangYuan/Deep-Learning-for-Microfluidic-Assisted-Caenorhabditis-elegans-Multi-parameter-Identification (accessed on 9 June 2023).
Micromachines 14 01339 g001 550
Figure 1. Schematic of deep learning for microfluidic-assisted C. elegans multi-parameter identification using YOLOv7.
Figure 1. Schematic of deep learning for microfluidic-assisted C. elegans multi-parameter identification using YOLOv7.
Micromachines 14 01339 g001
Micromachines 14 01339 g002 550
Figure 2. Numbers of annotated GFP and WT in the training, testing, validation, and external validation sets.
Figure 2. Numbers of annotated GFP and WT in the training, testing, validation, and external validation sets.
Micromachines 14 01339 g002
Micromachines 14 01339 g003 550
Figure 3. The overall structure is generic YOLO architecture. The backbone extracts the feature from the image. The neck enhances the feature map by combining different scales. The head provides the bounding box for detection. SPPCSPC, spatial pyramid pooling convolutional set pooling convolutional; Conv, convolution; Conv2D, two-dimensional convolutional layer; RepConv, reparameterized convolution; UpSampling2D, two-dimensional up-sampling layer; Concat, concatenate; CBS, Conv2D_BN_SiLU (convolutional standardization activation function); MCB, Multi_Concat_Block (multi-branch stacking module); TB, Transition_Block (transition block).
Figure 3. The overall structure is generic YOLO architecture. The backbone extracts the feature from the image. The neck enhances the feature map by combining different scales. The head provides the bounding box for detection. SPPCSPC, spatial pyramid pooling convolutional set pooling convolutional; Conv, convolution; Conv2D, two-dimensional convolutional layer; RepConv, reparameterized convolution; UpSampling2D, two-dimensional up-sampling layer; Concat, concatenate; CBS, Conv2D_BN_SiLU (convolutional standardization activation function); MCB, Multi_Concat_Block (multi-branch stacking module); TB, Transition_Block (transition block).
Micromachines 14 01339 g003
Micromachines 14 01339 g004 550
Figure 4. Multi-branch stacking module (i.e., MCB of Figure 3). The module is designed to enhance the accuracy of multi-target detection. During training, a module is split into multiple branches of the same or different modules. Then, multiple branches are integrated into a single module by designing multi-scale feature extraction. Finally, a multi-branch stacking module is added to further improve the precision of object detection.
Figure 4. Multi-branch stacking module (i.e., MCB of Figure 3). The module is designed to enhance the accuracy of multi-target detection. During training, a module is split into multiple branches of the same or different modules. Then, multiple branches are integrated into a single module by designing multi-scale feature extraction. Finally, a multi-branch stacking module is added to further improve the precision of object detection.
Micromachines 14 01339 g004
Micromachines 14 01339 g005 550
Figure 5. The training procedure and the general process of using YOLOv7 to detect C. elegans during sorting in microfluidic chips. (a) Videos from published microfluidic-assisted sorting studies were searched and captured as images, which were annotated using LabelImg software. (b) These annotated images were used in the training process of YOLOv7. (c) The model detected WT and GFP C. elegans and synthesized the final recognition output size and speed.
Figure 5. The training procedure and the general process of using YOLOv7 to detect C. elegans during sorting in microfluidic chips. (a) Videos from published microfluidic-assisted sorting studies were searched and captured as images, which were annotated using LabelImg software. (b) These annotated images were used in the training process of YOLOv7. (c) The model detected WT and GFP C. elegans and synthesized the final recognition output size and speed.
Micromachines 14 01339 g005
Micromachines 14 01339 g006 550
Figure 6. The trained model detected WT and GFP C. elegans in input images and output the recognized size and speed.
Figure 6. The trained model detected WT and GFP C. elegans in input images and output the recognized size and speed.
Micromachines 14 01339 g006
Table
Table 1. Parameters used to train YOLOv7, YOLOv5, and Faster R-CNN.
Table 1. Parameters used to train YOLOv7, YOLOv5, and Faster R-CNN.
ParametersValues
Optimizersstochastic gradient descent
Learning rate1 × 10−4~1 × 10−2
Momentum0.937
Learning decay5 × 10−4
PretrainedMicrosoft Common Objects in Context dataset
Number of epochs300
Table
Table 2. Performance comparison between YOLOv7, YOLOv5, and Faster R-CNN.
Table 2. Performance comparison between YOLOv7, YOLOv5, and Faster R-CNN.
ModelBackboneAnnotation NumbermAP (%)AP of the WT (%)AP of the GFP (%)
Faster R-CNNResnet50393198.9399.6498.22
YOLOv5Darknet-53393199.0310098.06
YOLOv7Darknet-53393199.5610099.12
Table
Table 3. Comparison of different models in target detection (YOLOv7, YOLOv5, and Faster R-CNN).
Table 3. Comparison of different models in target detection (YOLOv7, YOLOv5, and Faster R-CNN).
Metrics
ModelSample TypeDetection Speed (FPS)mAP (%)Model SizeAP (%)F1Recall (%)Precision (%)Annotation Number
Faster R-CNNWT19.03298.9310899.640.8410071.882426
GFP98.220.8599.3574.761505
YOLOv5WT32.14699.0327.110011001002426
GFP98.060.9696.1395.511505
YOLOv7WT52.11299.5614210011001002426
GFP99.120.9796.1397.391505
Table
Table 4. Comparison of external validation results among different models (YOLOv7, YOLOv5, and Faster R-CNN).
Table 4. Comparison of external validation results among different models (YOLOv7, YOLOv5, and Faster R-CNN).
Metrics
ModelSample TypeDetection Speed (FPS)mAP (%)Model SizeAP (%)F1Recall (%)Precision (%)Annotation Number
Faster R-CNNWT19.55066.2210864.320.5410036.84154
GFP68.120.5678.9542.86164
YOLOv5WT32.43488.8327.194.720.7390.4861.29154
GFP82.930.7363.1685.71164
YOLOv7WT52.50894.2114298.940.9595.2495.24154
GFP89.470.8989.4789.47164
Table
Table 5. Identification of WT and GFP C. elegans size and speed.
Table 5. Identification of WT and GFP C. elegans size and speed.
Sample TypeSampleSize (μm2)Speed (μm/s)
WTWorm125,310.880.00
WTWorm223,094.720.00
WTWorm371,830.65451.71
WTWorm413,989.5161.05
WTWorm516,067.160.00
WTWorm622,249.08511.05
WTWorm716,823.92970.29
WTWorm812,188.880.00
WTWorm911,547.360.00
WTWorm1034,372.351400.80
WTWorm1116,271.280.00
WTWorm1237,507.72 68.51
WTWorm1356,537.7855.95
WTWorm1493,056.3677.43
WTWorm1576,269.7461.26
WTWorm1652,304.8063.24
WTWorm1748,196.1893.28
WTWorm1837,507.7285.85
WTWorm1956,537.7867.00
WTWorm2055,724.9146.52
WTWorm2152,304.8046.13
WTWorm2248,196.1898.84
WTWorm2362,866.91131.96
WTWorm2468,319.6370.41
GFPWorm2540,444.920.00
GFPWorm2667,731.392827.30
GFPWorm2794,394.171216.61
GFPWorm2873,016.642028.55
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhang, J.; Liu, S.; Yuan, H.; Yong, R.; Duan, S.; Li, Y.; Spencer, J.; Lim, E.G.; Yu, L.; Song, P. Deep Learning for Microfluidic-Assisted Caenorhabditis elegans Multi-Parameter Identification Using YOLOv7. Micromachines 2023, 14, 1339. https://doi.org/10.3390/mi14071339

AMA Style

Zhang J, Liu S, Yuan H, Yong R, Duan S, Li Y, Spencer J, Lim EG, Yu L, Song P. Deep Learning for Microfluidic-Assisted Caenorhabditis elegans Multi-Parameter Identification Using YOLOv7. Micromachines. 2023; 14(7):1339. https://doi.org/10.3390/mi14071339

Chicago/Turabian Style

Zhang, Jie, Shuhe Liu, Hang Yuan, Ruiqi Yong, Sixuan Duan, Yifan Li, Joseph Spencer, Eng Gee Lim, Limin Yu, and Pengfei Song. 2023. "Deep Learning for Microfluidic-Assisted Caenorhabditis elegans Multi-Parameter Identification Using YOLOv7" Micromachines 14, no. 7: 1339. https://doi.org/10.3390/mi14071339

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop