Multi-Level Training and Testing of CNN Models in Diagnosing Multi-Center COVID-19 and Pneumonia X-ray Images

Featured Application

Despite their reported high accuracy, a significant limitation of current AI-assisted COVID-19 diagnostic models is that they are often trained on datasets sourced from specific clinics or possessing a limited number of training images. This raises an important question: Will these models maintain high accuracy when deployed in other clinics where images might exhibit disparities? If accuracy does drop, to what extent can we expect this decline? Conversely, how much can accuracy be improved by augmenting the training dataset with new images? In this study, we evaluated the performances of four CNN models that were trained on incrementally augmented datasets and subsequently tested on images with decreasing similarities. Through multi-level testing, we assessed the models’ capacities for verification, interpolation, and extrapolation in the context of diagnosing COVID-19 and pneumonia using multi-center X-ray images. Compared to conventional one-round training, multi-round training offers a more comprehensive insight into a model’s learnability, robustness, and interpretability.

Abstract

This study aimed to address three questions in AI-assisted COVID-19 diagnostic systems: (1) How does a CNN model trained on one dataset perform on test datasets from disparate medical centers? (2) What accuracy gains can be achieved by enriching the training dataset with new images? (3) How can learned features elucidate classification results, and how do they vary among different models? To achieve these aims, four CNN models—AlexNet, ResNet-50, MobileNet, and VGG-19—were trained in five rounds by incrementally adding new images to a baseline training set comprising 11,538 chest X-ray images. In each round, the models were tested on four datasets with decreasing levels of image similarity. Notably, all models showed performance drops when tested on datasets containing outlier images or sourced from other clinics. In Round 1, 95.2~99.2% accuracy was achieved for the Level 1 testing dataset (i.e., from the same clinic but set apart for testing only), and 94.7~98.3% for Level 2 (i.e., from an external clinic but similar). However, model performance drastically decreased for Level 3 (i.e., outlier images with rotation or deformation), with the mean sensitivity plummeting from 99% to 36%. For the Level 4 testing dataset (i.e., from another clinic), accuracy decreased from 97% to 86%, and sensitivity from 99% to 67%. In Rounds 2 and 3, adding 25% and 50% of the outlier images to the training dataset improved the average Level-3 accuracy by 15% and 23% (i.e., from 56% to 71% to 83%). In Rounds 4 and 5, adding 25% and 50% of the external images increased the average Level-4 accuracy from 81% to 92% and 95%, respectively. Among the models, ResNet-50 demonstrated the most robust performance across the five-round training/testing phases, while VGG-19 persistently underperformed. Heatmaps and intermediate activation features showed visual correlations to COVID-19 and pneumonia X-ray manifestations but were insufficient to explicitly explain the classification. However, heatmaps and activation features at different rounds shed light on the progression of the models’ learning behavior.

1. Introduction

Over the past few years, there has been a surge in studies exploring deep learning techniques to diagnose COVID-19 and pneumonia. Systematic reviews of AI-enabled COVID-19 detection can be found in [1,2,3] in 2021, [4,5,6,7,8,9,10] in 2022, and [11,12,13] in 2023. While all reviews delved into deep learning neural networks and medical image databases, there were apparent shifts in focus, from the feasibility of deep learning to this aim and limited databases in 2021, to model performance comparison and database survey in 2022, and to model enhancement/development and real-world applications recently. Because COVID-19 medical images were still limited in 2021, prior studies often focused on demonstrating the feasibility of deep learning models in distinguishing COVID-19 cases from normal subjects or patients with other respiratory diseases, such as pneumonia [14,15,16]. These studies either used chest X-ray images or CT scans [17,18,19,20,21,22] and were designed as either binary or multi-class classifications [23,24]. In general, the model classification accuracy was higher using chest X-ray images (90%+) than using CT (~85%) and was higher (1–5%) in binary than multi-class classifications [10,25,26]. To remedy data shortages, various techniques of data augmentation were used, including scaling, rotation, shifting, and image filtering such as grayscale and Gaussian blur [27,28,29].

With the increasing availability of COVID-19 X-rays and CT scans since 2021, research has pivoted to comparing various machine learning models and developing new models tailored for COVID-19 detection. For instance, Garg et al. [30] considered 20 CNN models for COVID-19 detection, including EfficientNet-B5, DenseNet169, InceptionV3, ResNet-50, and VGG16, using 4173 CT images, and demonstrated that EfficientNet-B5 and ResNet-50 were persistently superior in accuracy and sensitivity, followed by DenseNet, EfficientNet, and Xception, while VGG-19 remained the lowest across models. However, mixed results were frequently reported by other studies. Chouat et al. [31] compared ResNet-50, InceptionV3, VGG-19, and Xception using CT scans and X-ray images and reported that VGG-19 performed best on CT scans (87% accuracy) and Xception performed best on X-ray images (98% accuracy). In another study using CT images [32], ResNet-50 had the highest accuracy (96.97%), followed by Xception (90.71%), Inception-v3 (89.38%), and VGG16 (87.12%). Considering that such variability was reflective of the inherent advantages and setbacks of CNN models, hybrid or ensemble models leveraging multiple networks have been explored. By combining Inception V3 with VGG16, Srinivas et al. [33] achieved a 98% accuracy of COVID-19 prediction using 243 X-ray images, which outperformed Inception V3, VGG16, ResNet-50, DenseNet121, and MobileNet when tested individually. Similarly, Wang et al. [34] integrated features extracted from Xception, MobileNetV2, and NasNetMobile and made the classification via a confidence fusion method.

A notable variance in prior studies was the reported classification accuracies, ranging broadly from 78% to 100%, presumably due to the size or quality of the training/testing datasets. For instance, in 2021, Karar et al. [35] reported 99.9% classification accuracy for VGG-19 and ResNet-50. However, only 263 original X-ray images (including 56 normal, 49 COVID-19, and 128 pneumonia) were used, which were increased to 3325 images for each category through augmentation techniques like flipping, rotation, and shifting. In 2022, using 4326 chest X-ray images, Kumar et al. [36] reported an accuracy of 100% for binary classification (normal vs. COVID-19) and 98.82% for multi-class classification (normal, COVID-19, pneumonia). On the other hand, 78% accuracy using VGG-19 and 4137 CT images was reported by Garg et al. [30].

Despite exhibiting high accuracies in training and validation, CNN-based transfer learning might underperform during the testing phase [37,38,39,40,41]. This could be due to overfitting, which is a common issue in medical image classification with a limited number of discriminatory image features [7,13,42,43,44]. Prominent CNN models typically encompass over ten layers, boasting 60+ million trained parameters, and are trained on expansive datasets like ImageJ, which houses 1000 categories. In contrast, medical images have limited features, and their differences are often imperceptible to the human eye. This could be why the features/filters/convolutional layers trained on ImageJ may differ from those of medical images, resulting in lower testing performance [45]. In transfer learning, the adapted filters may retain irrelevant features, contaminating the classification process.

In summary, researchers have achieved accuracy levels exceeding 95% for COVID-19 detection through the selection and refinement of appropriate CNN models. Yet, systems based on these CNN models were often developed using data specific to certain clinics and acquired using their unique imaging modalities. Most proposed models or systems have not been tested on other datasets. It is also noted that even though the COVID-19 datasets have become more available, most have not been validated and can be subject to mislabeling, noise, incompleteness, corruption, or low quality [46,47,48,49]. One relevant question is how these models or systems that have been trained on one training database will perform on other datasets that inevitably have certain disparities. This question is even more pronounced when the training dataset contains a limited number of images. In this scenario, will the high accuracy of trained models be maintained when tested on larger data or images from other clinics? If not, how much lower accuracy will be expected and/or tolerated? Will adding new images to the original training dataset always improve the model’s performance, and what level of improvement can be expected via re-training the model with extended data? To achieve a desirable classification accuracy, how many new images should be added to the original training dataset?

The objective of this study was to evaluate the performance of CNN models in diagnosing COVID-19 and pneumonia in the lungs based on X-ray images from different data sources, as well as to assess the performance variation from multi-level training. Specific aims include the following:

• Comparing the three-class classification performances of four CNN models: AlexNet, ResNet-50, MobileNet, and VGG-19.
• For a given model, evaluating the model’s performance on X-ray images from different sources (i.e., inside and outside the training space).
• Quantifying the benefits of multi-round training with extended data on the model’s performance.
• Evaluating the model’s interpretability via heatmaps and intermediate activation features.

2. Methods

2.1. COVID-19 and Pneumonia Datasets from Multiple Sources

Chest X-ray images were selected from three sources, with each source encompassing the categories of normal, COVID-19, and pneumonia. The first dataset [50,51,52] contains 10,192 normal cases, 3616 COVID-19 positive cases, and 1345 viral pneumonia images. Among these, approximately 80% were used for training (11,538) and 20% for testing (3035), as shown in the left panel of Figure 1a. Both the training and testing datasets were divided in such a way that each split retained approximately the same proportion of samples for each class as in the original dataset (i.e., stratified random sampling). Among the 11,538 training images, 7791 are normal, 2712 are COVID-19, and 1030 are viral pneumonia, constituting a disproportionate distribution of 68%, 24%, and 9%, respectively. Likewise, among the 3035 testing images, 2032 are normal, 742 are COVID-19, and 261 are viral pneumonia. Three sample images typical of normal, COVID-19, and viral pneumonia cases are also shown below.

The second dataset [53] contains 6432 images. Like Dataset 1, stratified random sampling was implemented to set apart 20% (1288) as the level 2 test images, which comprised 855, 317, and 116 for normal, COVID-19, and pneumonia, respectively (Figure 1a). Multi-center X-ray images were tested to evaluate model applicability in external centers.

A subfolder named “Outliers” was created by collecting images from the first and second datasets that are apparently distorted or irregular. Some radiographic images exhibited features such as rotation, distortion, differing magnifications, and variations in contrast and brightness, as shown in the lower panel of Figure 1a. These outlier images were excluded from the Round 1 training dataset. Instead, they were trained at designated proportions (25% and 50%) in Rounds 2 and 3. The “Outliers” dataset consisted of 247 images and was also used to test CNN models as level 3 test samples. It contained 104 normal, 107 COVID-19, and 36 pneumonia cases. Considering that these images have a lower level of similarity to the training images, lower classification accuracies are expected on this test dataset in comparison to the level 1 and level 2 testing datasets.

The third dataset [54] contains 7135 X-ray images. Our initial study indicated that CNN models trained on Dataset 1 consistently performed well on Dataset 2 but poorly on Dataset 3. We thus inferred that the images in Dataset 3 had higher levels of differences. In this study, around 10% (730 images) of Dataset 3 was designated for testing via stratified random sampling, which was used as the Level 4 testing dataset. It comprised 234 normal, 105 COVID-19, and 390 pneumonia cases (right panel, Figure 1a). This dataset functioned similarly to dataset 2 (or Level 2) and was used to substantiate the statistical significance by testing images from an external medical center.

2.2. Selection of CNN Models

The selection of CNN models depended on the training purpose and the model’s specificities. The purpose of this study was to develop a CNN-based automated diagnostic system with high performance for COVID-19 and pneumonia patients based on chest X-ray images. Relevant criteria included the following: high classification accuracy, high sensitivity and specificity, robust performance across centers, the ability to learn continuously, and the ability to identify discriminative features among categories. Features learned are often abstract, beyond human comprehension or interpolation. However, sometimes they can capture some inherent differences that are consistent with our perceptions, which can be useful in understanding either the etiology or manifestations of a specific disease. Ideally, only defining features are learned from the training dataset; subsequent addition of highly similar images will not bring new features to the CNN model and thus can only marginally improve the CNN model. Only new images with new features introduced to the model will see extra features to refine the classification of the test dataset.

Four convolutional neural network (CNN) models were selected for this study: AlexNet, ResNet-50, MobileNet, and VGG-19. AlexNet and ResNet-50 were selected because they were the 2012 and 2015 winners of the ImageNet competition, respectively [32,33]. AlexNet was groundbreaking in its use of GPUs for training deep neural networks, while ResNet-50 introduced residual connections between different layers to improve gradient flow and enable the training of even deeper neural networks [34,35]. MobileNet was chosen for its simpler architecture and smaller computational requirements [36,37,38,39]. It will be desirable to run a computer-aided diagnostic (CAD) system on a personal computer or even a smartphone, provided it can achieve sufficient diagnostic accuracy. VGG-19 was selected for its simplicity, which uses only 3 × 3 convolutional layers stacked on top of each other in increasing depths. One drawback of VGG-19 is its requirement for more memory than the other three models, primarily due to its fully connected layers.

To determine whether the architecture of a pre-trained network is suitable for a test dataset, the following steps were taken. First, the network was trained and tested on the same dataset, providing an initial evaluation of the model’s performance on one testing dataset. In this step, various metrics are evaluated, including accuracy, AUC, specificity, sensitivity, precision, and ROC, to identify areas for improvement (Figure 1c). Second, the pre-trained network was fine-tuned based on the identified areas for improvement by adjusting the hyperparameters or replacing some of its layers to improve its performance. After fine-tuning, the pre-trained network was re-evaluated on the test dataset until its performance reached a threshold classification accuracy.

The fine-tuned network model was further tested on additional datasets from the same medical center that the model had never seen before (for instance, newly acquired images). This step evaluates the pre-trained model’s interpolation capacity.

In the following step, the model was tested on datasets from other medical centers that belong to the same category (e.g., COVID-19 images, but the model has never seen them before). This step evaluates the pre-trained model’s extrapolation capacity. Even though the X-ray images acquired at any medical center share inherent similarities among COVID-19 patients, differences among COVID-19 images can exist across medical centers, and the performance of the pre-trained model can be affected by such differences.

In the next step, the model was re-trained by adding new images to the training dataset and re-tested. In this step, two metrics are evaluated: (1) the improvement in performance, and (2) the minimal number of images needed to reach a satisfactory classification accuracy. In doing so, it is hoped to provide a practical example of the CNN-based system when applied in medical centers other than the one where the system was originally trained.

2.3. Study Design for Multi-Level CNN Model Training and Testing

The training protocol was conceived to have five rounds, and in each round, testing was conducted at four levels on images differing in quality or similarity (as detailed in

Table 1. Five-round (R1–5) training and four-level testing protocol to evaluate models’ capacity for interpolation, extrapolation, and continuous learning. These procedures will be tested on four models (AlexNet, ResNet-50, MobileNet, and VGG-19) for three-class classifications (normal, COVID-19, and pneumonia).

	Training	Testing
		Level 1	Level 2	Level 3	Level 4
R1	Base	Dataset 1	Dataset 2	Outliers	Dataset 3
R2	Base, 25%L3 *	Dataset 1	Dataset 2	Outliers	Dataset 3
R3	Base, 50%L3	Dataset 1	Dataset 2	Outliers	Dataset 3
R4	Base, 50%L3, 25%L4 *	Dataset 1	Dataset 2	Outliers	Dataset 3
R5	Base, 50%L3, 50%L4	Dataset 1	Dataset 2	Outliers	Dataset 3

* L3, L4: Level 3, Level 4.

). In each test, a three-class classification (normal, COVID-19, and pneumonia) was performed. By training one model through several rounds with augmented datasets and testing its performance on datasets with decreasing similarities, it was aimed to (1) select the optimal CNN model, (2) test the model’s ability for interpolation and extrapolation, and (3) test the model’s ability to learn from new data.

Round 1, referred to as the “baseline training”, incorporated 11,538 radiographic images (baseline, or “Base” for short in Table 1). It was aimed at validating the CNN models (i.e., Level 1) as well as evaluating whether the models can classify new samples that are either similar (Level 2), outliers (Level 3), or have higher levels of dissimilarity (Level 4). In doing so, a preliminary evaluation of the model’s interpolation and extrapolation capacity was obtained, based on which further training and testing could be designed.

To evaluate the effects of re-training on model performance, five rounds of training were performed in this study with incrementally extended data. The selection of added images was determined by the testing results in Round 1. In Round 2, 25% of the outlier (Level 3) images were added in the hope that the CNN model could learn the inherent features of those outliers and thus improve the classification accuracy of these outlier images. In Round 3, 50% of the outlier images were added. In comparison to Round 2, this round aimed to quantitatively evaluate the performance improvement as a function of added images.

Round 2 and Round 3 training exemplified the processes of model development and optimization within one medical center based on its own data. In contrast, Rounds 4 and 5 training represented scenarios where the automated diagnostic system was used in other medical centers, and when generating sub-optimal performances, it was further improved by including new images to re-train the system. Similarly, the two-round training (Rounds 4 and 5) was designed to quantify performance improvements vs. added images. Round 4 included 25% of dataset 3 training images in addition to the Base and 50% outlier, while Round 5 included 50% of dataset 3 training images (Figure 1b, lowest row).

To evaluate the network classification performance, various indices were quantified, including accuracy, sensitivity, specificity, precision, AUC (area under the curve), and ROC (receiver operating characteristic) curve. For three-class classification (i.e., 1, 2, 3), the performance matrices were extended from their binary counterparts. Based on a three-class confusion matrix (blue dashed rectangle) in Figure 1c, the accuracy is the ratio of correct predictions over all predictions (Figure 1c, left bottom). The category-wise matrices (specificity, sensitivity, and precision) were calculated using the equations listed to the right of Figure 1c. For the Receiver Operating Characteristic (ROC) curve, the One-vs-Rest approach (OvR) was used, e.g., COVID vs. (normal plus pneumonia). In this study, all category-wise matrices were presented for COVID unless stated otherwise.

The CNN model training/testing was conducted using an AMD Ryzen 3960X 24-Core workstation with 3.79 GHz processors, 256 GB RAM, and 24 GB GPU (PNY NVIDIA GeForce RTX 3090). For the approximately 12,000 images used in this study, one round of training for one CNN model took approximately 4–7 h.

3. Results

3.1. Round 1 Training and Testing (Model Development and Verification)

Table 2. Round 1 performance comparison among models (AlexNet, ResNet-50, MobileNet, and VGG-19) that were trained on Base and tested on samples with varying similarities (Level 1–4) for three-class classifications (normal, COVID-19, pneumonia).

Network	Round 1 (%)	Level 1	Level 2	Level 3	Level 4
AlexNet	Accuracy	98.91	97.93	57.44	84.11
	AUC	99.94	99.95	78.52	99.47
	Specificity	97.71	95.75	74.10	97.78
	Sensitivity	99.51	99.19	34.95	55.13
	Precision	98.88	97.60	50.00	92.14
ResNet-50	Accuracy	99.08	97.93	58.26	87.40
	AUC	99.97	99.91	76.46	98.22
	Specificity	97.21	94.34	72.66	91.94
	Sensitivity	100.00	100.00	38.33	77.78
	Precision	98.64	96.85	51.28	81.98
MobileNet	Accuracy	99.24	98.28	52.89	86.30
	AUC	99.93	99.98	83.12	99.40
	Specificity	97.91	97.17	71.22	95.56
	Sensitivity	99.90	98.92	28.16	66.67
	Precision	98.98	98.38	42.03	87.64
VGG-19	Accuracy	95.22	94.66	55.79	84.66
	AUC	99.24	99.15	73.31	98.93
	Specificity	91.43	90.09	64.75	92.74
	Sensitivity	97.10	97.29	43.69	67.52
	Precision	95.82	94.47	47.87	81.44

summarizes the model performance in Round 1 of four deep neural networks (i.e., AlexNet, ResNet-50, MobileNet, and VGG-19) on test datasets from various sources and with four levels of image similarity.

At Level 1, all models, except for VGG-19, showed exceptional performance, achieving accuracy rates of over 98%. The AUC, an indicator of a model’s capacity to distinguish between classes, was above 99% for all models. ResNet-50 led this metric with an AUC of 99.97%, followed closely by AlexNet, MobileNet, and VGG-19 with 99.94%, 99.93%, and 99.24%, respectively.

At Level 2, with X-ray images from a different source, the models maintained high performance, albeit with a slight decrease in most metrics. Although the sensitivity remained high across all models, the specificity showed a minor decrease compared to Level 1, yet it remained sufficiently high, underlining an efficient ability to correctly identify negative cases. Among the four models considered, VGG-19 exhibited the lowest specificity (~90%) when tested on both Level 1 and Level 2 images. The other three models all had a specificity of 94% or higher.

However, progressing from Level 2 to Level 3, which comprised outliers from datasets 1 and 2, there was a drastic decline in performance in all models, most noticeably in sensitivity and precision metrics (Figure 2). This decrease was likely attributed to the increased complexity or alterations in image presentations at Level 3. In particular, the sensitivity of MobileNet decreased from 97%+ to 28% (Table 2). The large drop in sensitivity suggested a diminishing ability of the models to correctly identify positive cases when tested on Level 3 or the outlier images. Similarly, a decline in precision from 97% at Level 2 to 48% on average (Figure 2) at Level 3 implied a reduced proportion of correct positive identifications.

At Level 4, which represented images from an external medical facility, all models showed a decline in classification accuracy compared to Level 2, e.g., from 97% to 86% on average. Although the AUC values remained high for all four models (~99%), both sensitivity and precision experienced a substantial drop. On average, from Level 2 to Level 4, sensitivity decreased from 99% to 67%, and precision decreased from 97% to 86%, as shown in Figure 2. This trend indicated a reduced efficiency of the models in accurately identifying true positive cases and in the precision of positive identifications in this setting.

3.2. Round 2 Training and Testing (Model Improvement by Adding 25% Outliers)

Table 3. Round 2 performance comparison among models (AlexNet, ResNet-50, MobileNet, and VGG-19) that were trained on Base and tested on samples with varying similarities (Level 1–4) for three-class classifications (normal, COVID-19, pneumonia).

Network	Round 2 (%)	Level 1	Level 2	Level 3	Level 4
AlexNet	Accuracy	98.15	98.28	71.90	80.55
	AUC	99.87	99.94	91.72	99.26
	Specificity	97.05	96.70	75.54	98.19
	Sensitivity	98.68	99.19	66.99	43.16
	Precision	98.58	98.12	66.99	91.82
ResNet-50	Accuracy	99.54	98.45	77.69	86.71
	AUC	99.89	99.96	92.93	99.91
	Specificity	99.00	97.17	81.29	97.18
	Sensitivity	99.80	99.19	72.82	64.53
	Precision	99.51	98.39	74.26	91.52
MobileNet	Accuracy	98.76	98.80	71.07	83.84
	AUC	99.88	99.99	91.91	98.90
	Specificity	98.11	98.11	75.54	96.98
	Sensitivity	99.08	99.19	65.05	55.98
	Precision	99.08	98.92	66.34	89.73
VGG-19	Accuracy	98.62	94.66	62.40	82.19
	AUC	99.92	99.42	83.22	98.73
	Specificity	98.31	91.51	61.87	94.15
	Sensitivity	98.77	96.48	63.11	56.84
	Precision	99.16	95.19	55.08	82.10

shows the results of the second round of training (Round 2) using the base images supplemented with 25% outliers. A comparison between R2 and other rounds is shown in Figure 3. Considering Level 1 and Level 2 testing results, all models demonstrated high performance with accuracy exceeding 98% and AUC exceeding 99.8% (except VGG-19 at Level 2 with a 95% accuracy). Compared to Round 1, the sensitivity of all models remained high (~99%) despite a minor decrease (<1%); meanwhile, both the specificity and precision increased slightly (~1%).

At Level 3, all considered models exhibited marked performance improvements, suggesting that the addition of 25% outliers, equating to 62 randomly selected images, effectively addressed the shortfall in outlier-related features. Notably, the classification accuracy increased from 56% to 71% on average, while the sensitivity jumped from 36% to 67% (Figure 3a). This sharp rise in sensitivity highlighted two factors: (1) The Round 1 training dataset (i.e., Base) shared minimal commonality with outlier discriminative features, resulting in notably low accuracy and sensitivity, and (2) even adding a small number (62) of relevant images to a large training dataset (Base, 11,538 images) could significantly improve model performance, particularly the sensitivity. Other performance metrics also improved substantially, with AUC increasing from 78% to 90% and precision from 48% to 66%. There was a slight increase in specificity, from 71% to 74%, likely due to the significant imbalance in the distribution of outliers compared to regular images in the training datasets (62 vs. 11,538).

Round 2 model performance at Level 4 testing (external center) was marked by a slight decrease in classification accuracy (i.e., from 86% to 83%) but a large decrease in sensitivity (i.e., from 67% to 55%, Figure 3b). On the other hand, the specificity increased slightly, from 95% to 97% (Figure 3b). The AUC and precision metrics displayed mixed trends and varying magnitudes (Table 2).

3.3. Round 3 Training and Testing (Model Refinement by Adding 50% Outliers)

Further adding outliers (an additional 25%, or 62 images) to the Round-2 training dataset further improved the model performance on Level 3 testing. In comparison to Round 2, the accuracy increased from 71% to 83% and the sensitivity from 67% to 86% on average for all models considered, as listed in

Table 4. Round 3 performance comparison among models (AlexNet, ResNet-50, MobileNet, and VGG-19) that were trained on Base and tested on samples with varying similarities (Level 1–4) for three-class classifications (normal, COVID-19, pneumonia).

Network	Round 3 (%)	Level 1	Level 2	Level 3	Level 4
AlexNet	Accuracy	99.11	98.80	85.54	82.05
	AUC	99.94	99.98	97.92	99.71
	Specificity	98.01	98.11	84.89	97.38
	Sensitivity	99.66	99.19	86.41	49.57
	Precision	99.02	98.92	80.91	89.92
ResNet-50	Accuracy	98.45	97.76	88.02	77.12
	AUC	99.90	99.89	96.82	99.40
	Specificity	97.31	95.75	84.89	76.61
	Sensitivity	99.02	98.92	92.23	78.21
	Precision	98.68	97.59	81.90	61.20
MobileNet	Accuracy	98.91	98.80	84.71	82.47
	AUC	99.89	100.00	96.91	99.78
	Specificity	97.51	98.58	82.73	97.78
	Sensitivity	99.61	98.92	87.38	50.00
	Precision	98.78	99.18	78.95	91.41
VGG-19	Accuracy	95.16	93.12	73.55	81.51
	AUC	99.15	99.23	92.35	99.09
	Specificity	89.43	84.91	69.78	95.56
	Sensitivity	97.98	97.83	78.64	51.71
	Precision	94.95	91.86	65.85	84.62

and Figure 3a. The other three metrics also improved, with the AUC increasing from 90% to 96%, specificity from 74% to 81%, and precision from 66% to 77% on average (Table 4 and Figure 3a). This indicates that there was still room for model improvement after the addition of 164 images to the Base training dataset. By comparison, the model performance remained relatively unchanged at Level 1 and Level 2, partially because of the small proportion of images added to the Base training dataset. Insignificant changes between Round 2 and Round 3 were also observed at Level 4 (Figure 3b), due to the weak correlation between the Outliers and Level 4 dataset.

3.4. Round 4 and 5 Training and Testing

For the Level 4 testing dataset, model classification accuracies were persistently observed to remain around 80% in Rounds 1–3. This was reasonable considering that the Rounds 1–3 models were trained predominately on Dataset 1, which might miss certain features inherent in Level 4 images (Dataset 3). This underscores the fact that simply adding irrelevant images to the training set does not enhance model performance, reinforcing the pivotal importance of data quality and relevance in model training. In Rounds 4 and 5, 25% and 50% of the Level 4 images (730) were added to the Round-3 training dataset, respectively. The performance metrics are listed in

Table 5. Round 4 performance comparison among models (AlexNet, ResNet-50, MobileNet, and VGG-19) that were trained on Base and tested on samples with varying similarities (Level 1–4) for three-class classifications (normal, COVID-19, pneumonia).

Network	Round 4 (%)	Level 1	Level 2	Level 3	Level 4
AlexNet	Accuracy	98.91	98.45	85.95	90.96
	AUC	99.95	99.98	97.73	99.89
	Specificity	98.01	97.17	86.33	98.59
	Sensitivity	99.36	99.19	85.44	74.79
	Precision	99.02	98.39	82.24	96.15
ResNet-50	Accuracy	98.85	98.11	86.36	93.01
	AUC	99.93	99.87	96.00	99.98
	Specificity	97.51	96.23	84.17	90.73
	Sensitivity	99.51	99.19	89.32	97.86
	Precision	98.78	97.86	80.70	83.27
MobileNet	Accuracy	99.08	98.28	85.54	93.15
	AUC	99.91	99.97	97.33	99.87
	Specificity	98.21	97.17	85.61	98.79
	Sensitivity	99.51	98.92	85.44	81.20
	Precision	99.12	98.38	81.48	96.94
VGG-19	Accuracy	94.86	94.32	71.49	90.55
	AUC	98.94	99.01	91.01	99.70
	Specificity	88.53	88.68	67.63	95.36
	Sensitivity	97.98	97.56	76.70	80.34
	Precision	94.54	93.75	63.71	89.10

and

Table 6. Round 5 performance comparison among models (AlexNet, ResNet-50, MobileNet, and VGG-19) that were trained on Base and tested on samples with varying similarities (Level 1–4) for three-class classifications (normal, COVID-19, pneumonia).

Network	Round 5 (%)	Level 1	Level 2	Level 3	Level 4
AlexNet	Accuracy	98.85	98.28	84.30	94.25
	AUC	99.96	99.90	97.66	99.91
	Specificity	98.01	97.64	84.89	98.99
	Sensitivity	99.26	98.64	83.50	84.19
	Precision	99.02	98.64	80.37	97.52
ResNet-50	Accuracy	98.45	98.45	83.06	95.89
	AUC	99.91	99.96	95.97	99.92
	Specificity	97.61	97.17	83.45	94.56
	Sensitivity	98.87	99.19	82.52	98.72
	Precision	98.82	98.39	78.70	89.53
MobileNet	Accuracy	99.04	98.97	86.78	96.16
	AUC	99.88	100.00	97.09	99.98
	Specificity	97.81	97.64	84.17	98.99
	Sensitivity	99.66	99.73	90.29	90.17
	Precision	98.93	98.66	80.87	97.69
VGG-19	Accuracy	95.02	93.46	73.55	91.92
	AUC	99.19	99.31	91.42	99.93
	Specificity	88.63	84.43	71.94	96.57
	Sensitivity	98.18	98.64	75.73	82.05
	Precision	94.59	91.69	66.67	91.87

.

From Table 5, it was noted that the classification metrics at Levels 1–3 changed insignificantly (e.g., 0.1–2%) between Round 4 and Round 3 for all models considered. The metric variations were also mixed in trend. Sensitivity was the only exception, experiencing a decrease across all models. This suggests that introducing new, non-relevant features or images can undermine the model’s capacity to recognize pertinent features.

At Level 4, adding 25% Level-4 images to the Round-3 training dataset resulted in a substantial increase in all performance metrics. Compared to Round 3, the sensitivity increased by 27% (from 57% to 84%), accuracy by 11% (from 81% to 92%), precision by 9% (from 82% to 91%), specificity by 4% (from 92% to 96%), and AUC by 0.4% (from 99.5% to 99.9%), as shown in both Table 5 and Figure 3b. Note that an average classification accuracy of 92% is deemed satisfactory in many clinical applications, underscoring the benefits of multi-round training with extended data, even when exposed to a limited set of new images (in this case, 25% of 730 or 183 images).

Table 6 shows the Round-5 training/testing results after adding another 25% of Level-4 images to the Round-4 training dataset. The classification accuracy tested on the Level-4 images reached 95%, up from 92% in Round 4 (Figure 3b). The sensitivity showed the largest magnitude of increase, i.e., 5% (from 84% to 89%). On the other hand, the specificity increased from 96% to 97% (1%) and the precision from 91% to 94% (3%), as shown in Figure 3b.

At Level 3, small variations with mixed trends were observed between Round 4 and 5 trained models. Overall, adding training images that were irrelevant to Level 3 led to a slight decrease in model performance (0.4% on average, Table 6). Comparing the four models, much lower precision and accuracy were observed for VGG-19 than for other models (i.e., 67% vs. 80 ± 1% and 74% vs. 85 ± 1.5%, respectively), as shown in Figure 4. This observation persisted in Round 4 (64% vs. 81 ± 0.6% and 71% vs. 86 ± 0.3%), Round 3 (66% vs. 81 ± 1.2% and 74% vs. 86 ± 1.4%), and Round 2 (55% vs. 69 ± 3.6% and 62% vs. 74 ± 2.9%), even though VGG-19 had an equivalent Level-3 precision and accuracy in Round 1 (48% vs. 48 ± 4.1% and 56% vs. 56 ± 2.4%, Figure 4). Thus, based on the datasets in this study, VGG-19 exhibited a diminished response to re-training with extended data, compared to the other three models.

3.5. Model Selection

ROC (receiver operating characteristic) curves are displayed for the training phase of Rounds 1, 2, and 4. The ROC curves at Levels 1 and 2 are near perfect and thus are not presented. At Level 3 (upper panel, Figure 5), for the Outliers testing dataset, re-training with extended data moved all curves toward the top-left corner of the graph, indicating improved discriminatory power for all models considered (Figure 5a–c). Among them, VGG-19 exhibited the lowest improvement in all training phases. The other three models showed insignificant discrepancies in Round 4 (Figure 5c); however, in Round 2, ResNet-50 showed the best performance (learnability), with a much lower threshold of false positive rate for a 95% detection rate than AlexNet and MobileNet.

At Level 4 (lower panel, Figure 5), for external testing images, the ROC curves were already skewed towards the upper-left corner. Model performances constantly increased from Round 1 to Round 4, and ResNet-50 remained the optimal one among the four.

The ROC curves of ResNet-50 are compared across Rounds 1–5 in Figure 6 at Level 3 and Level 4. Note that in Round 2 and Round 3, the training dataset was extended by adding 25% and 50% outlier images (Level 3). As a result, the detection rate on the Level 3 dataset increased drastically from Round 1 to Round 2, as well as from Round 2 to Round 3, but at a lower rate (Figure 6a). In Rounds 4 and 5, adding external source (Level 4) images did not improve ResNet-50′s detection rate on the Level 3 dataset, but rather decreased it slightly (Figure 6a). At Level 4 (i.e., dataset from external centers), very high detection rates were achieved by adding external but relevant training images, as evidenced by the ROC curves that were highly skewed towards the upper-left corner. As a result, the ResNet-50 model trained in Round 5 was more accurate and had wider applicability than the ResNet-50 model trained in Round 1.

To evaluate the multi-round training effects on category-wise performance (i.e., normal, COVID-19, and pneumonia), the misclassification rate at Level 4 by ResNet-50 was further broken down into individual groups in Round 1 and Round 5, as shown in Figure 7a. Note that in Round 1, the Level-4 ResNet-50 accuracy was 87.4% (Table 2), which resulted from individual accuracies of 78.8% for normal, 83.0% for COVID-19, and 94.4% for pneumonia. For each category, the misclassification rates for one of the other two categories were also calculated. For instance, a 22.2% misclassification rate (100–78.8%) of normal cases included 2.1% of cases inaccurately predicted as COVID-19 (brown color) and 20.1% as pneumonia (grey color), as shown in Figure 7a. Similarly, a 17.7% misclassification rate of COVID-19 all came from their incorrect prediction as normal, while a 5.6% misclassification rate of pneumonia included 5.3% incorrectly predicted as normal and the remaining 0.3% as COVID-19 (Figure 7a).

In Round 5, the misclassifications of normal as COVID-19 and pneumonia were greatly reduced, improving the model’s specificity (right vs. left panels, Figure 7a). A similarly significant reduction was also observed in the misclassification of COVID-19 as normal, thus improving the model’s sensitivity. However, the misclassification of COVID-19 as pneumonia increased (i.e., from 0% to 1.9%). The misclassification of pneumonia as normal also increased slightly (i.e., from 5.1% to 6.2%, Figure 7a), which would partially reduce the model’s sensitivity.

For comparison purposes, the VGG-19 misclassifications in individual categories are shown in Figure 7b in the left and right panels for Round 1 and Round 5, respectively. Note the same range in the y-coordinate (0–25%) in these two panels. Although there were noteworthy decreases in misclassification, the scale of these reductions was evidently less significant compared to ResNet-50, especially in the mispredictions from normal to pneumonia and from COVID-19 to normal. This implied that VGG-19 had a lower learning capacity compared to ResNet-50 for the dataset in this study.

Interestingly, the misclassification rate between COVID-19 and pneumonia is very low, with the highest rate of 1.9% for ResNet-50 (COVID-19-to-pneumonia in Round 5) and 1.8% for VGG-19 (pneumonia-to-COVID-19 in Round 1), as shown in Figure 7a,b. This low mutual misclassification indicated that discriminatory features between these categories might have been adequately identified by both CNN models.

3.6. Model Visualization

Heatmaps provide a visualization of which parts of an image are important for classification decisions. These parts can be correlated to the clinical or empirical observations that physicians rely on to reach their diagnostic decisions. Figure 8a shows the heatmaps in terms of occlusion sensitivity in the sample images of normal, COVID-19, and pneumonia in Round 1. Hotspots are those that, when occluded, will cause a significant drop in the model’s prediction. Interestingly, the heatmaps differed remarkably among models, even for an identical image, indicating very different discriminatory features used in different CNN models. Considering the normal image, no hotspots were predicted within the lung by all models except VGG-19. For COVID-19 (second row), scattered zones of intermediate sensitivities (opacities) were predicted within the lungs. By contrast, sporadic hotspots were observed in the pneumonia image. In particular, a large zone with intermediate-to-high sensitivities was employed in the lower lobes by ResNet-50.

Figure 8b shows the occlusion sensitivity using ResNet-50 in Round 5. After four more rounds of training by exposing them to an incrementally enriched dataset, the heatmaps in Round 5 appeared more indicative and intuitive than in Round 1 with respect to the clinical manifestations, which exhibited a clear lung in the normal case, dispersed opacities and occasional hotspots in COVID-19, and a large region of hotspots in the lower lobes in pneumonia. The improved interpretability from Round 1 to Round 5 underscored the potential benefits of multi-round training on model performance.

Next, we examined the impact of multi-round training (Rounds 1–5) on the Grad-CAM heatmap predicted by ResNet-50 (Figure 9). Grad-CAM, short for Gradient-weighted Class Activation Mapping, computed the gradients of the model’s class scores with respect to the final convolutional layer’s feature maps and only kept features with positive contributions. From the heatmap variations across Rounds 1–5, we could see the progression of ResNet-50’s learning behavior by gradually including more diverse images. It was noted that the three images were all correctly predicted. Like the occlusion sensitivity in Figure 8, the Grad-CAM heatmap of the normal image was clear in the lung in all five rounds (the upper panels, Figure 9a–e), which was consistent with the final prediction for the normal class. Considering the COVID-19 image, hotspots of Grad-CAM were observed mainly in the middle and upper lungs. However, the intensity and location of these hotspots varied from Round 1 to Round 5, and this variation did not show a regular pattern in this example. Thus, explicit correlation from heatmaps to clinical manifestations of the disease needs caution, even though it is highly desirable to detect the disease from heatmaps and learned features or to use heatmaps and learned features to better understand the disease.

The intermediate activation features at selected layers in residual Block 1 of ResNet-50 are shown in Figure 10. A diagram of the main layers in Block 1 is shown in Figure 10a, which has two branches (i.e., shortcut branch and main branch) and represents the key innovation in ResNet-50, which allows activations to bypass one or more layers. Here “bn_conv1” stands for “batch normalization for convolution layer 1”. The same three source images were passed to the model (Figure 10b), and Figure 10c shows sample extracted features at the first convolution layer “conv1”, followed by batch normalization (Bn_conv, Figure 10d), Max_pooling (Figure 10e), Bn2a_brach2a (Figure 10f), Bn2a_brach2b (Figure 10g), and the Block 1 activation layer “Add_1” (Figure 10h). Different features were extracted from the input images, e.g., skeleton, lung tissue, texture, brightness, edges, etc. As the layer went deeper, the features became less recognizable with refined scales. Note that in each layer in Block 1, there were either 8 × 8 or 16 × 16 features, and only 2 × 2 was shown here. In Max_pooling (Figure 10e), the dark color shows less activated features, and the light color shows highly activated features. At the end layer of residual Block 1 (Add_1), the processed features of the input images through the main and shortcut branches were added, which appeared to bear more visual similarity to the input images than features in the intermediate layers.

4. Discussion

4.1. Model Selection

Among the four models considered herein (AlexNet, ResNet-50, MobileNet, and VGG-19), all models gave satisfactory classification performance in Round 5 on the three databases (91.9–99.0%). Thus, these four models are all good candidates for developing AI-assisted COVID-19 diagnostic systems, as also suggested by many previous studies. Among them, ResNet-50 consistently had higher overall performance than other networks across the five training stages and four levels of testing. It also outperformed other networks in learning new features when progressively exposed to more diverse training images, as evidenced by the more upper-left skewed ROC curves in Figure 5 when compared to other models and the quickly improved ROC curves across Rounds 1–5 in Figure 6.

VGG-19 showed the lowest performance in this study. However, it still had a classification accuracy of 95.0% on the first database (Level 1), 93.5% on the second database (Level 2), and 91.9% on the third database (Level 4). It performed poorly on the outlier images collected from the first and second databases, even after being trained on 50% of them. The classification accuracy in Round 5 was 73.6%, compared to 83.1–86.8% achieved using other networks. Considering its persistent underperformance in five rounds of training and four levels of testing, we recommended ResNet-50 or MobileNet over VGG-19 for the future development of automatic diagnostic systems using X-ray images. MobileNet was recommended because of its simpler architecture (thus fewer computational requirements), overall satisfactory classification accuracy, robust performance, and good learnability.

4.2. Dataset Effects

In this study, three datasets of X-ray images for normal, COVID-19, and pneumonia were utilized. We demonstrated that, when using the same datasets for training and testing, high classification accuracy (98.1 ± 1.6%) and sensitivity (99.1 ± 1.2%) were achieved even for the testing images that had been set apart in a separate folder (Level 1), as shown in Figure 2 and Table 2. Furthermore, even for an external dataset (Level 2), impressive accuracy (97.2 ± 1.5%) and sensitivity (98.6 ± 1.0%) were achieved, indicating the high quality of both datasets as well as the image resemblance between these two datasets.

The Level 3 dataset contained outlier images collected from the first and second databases, which exhibited evident deformities (rotation, scaling, shifting, incompleteness) and color variations. These images presumably came from data augmentations when the database was small. In this study, we observed drastic declines in model performance when tested on these outlier images—accuracy dropped from 98.1% at Level 1 to 56.1%, sensitivity from 99.1% to 36.3%, and precision from 98.1% to 47.8%. Note that CNN models in Round 1 had not seen these outlier images yet, even though the outlier images were modified from the training images. The drastic drop in performance indicated that features of the outlier images were missing in Round-1-trained models, indicating that the traditional data augmentation techniques would introduce extraneous features and undermine the model development using chest X-rays. Even though data argumentation techniques have been useful when the COVID-19 images were limited, training unmodified images is recommended when COVID-19 X-rays become more available.

The Level 4 dataset in this study comprised external images that displayed certain disparities from the training images, as evidenced by the moderately reduced classification metrics from Level 1 to Level 4 (i.e., accuracy from 98.1% to 85.6%, sensitivity from 99.1% to 66.8%, Figure 2 and Table 2). Even within the same modality (X-ray), variations could arise due to differences in equipment and operational procedures. The heterogeneity in image quality could introduce potential biases in model training and superfluous misclassifications in testing [55]. In addition, images are susceptible to mislabeling and corruption. Thus, standardization and validation of image databases and repositories, which are lacking in most public COVID-19 X-ray datasets, are needed to avoid future complications.

The training data were imbalanced in this study. The baseline training images (11,538) include 7791 for normal, 2717 for COVID-19, and 1030 for pneumonia. This might explain the high misclassification rates in normal vs. COVID-19 and normal vs. pneumonia (Figure 7). By contrast, the classification for COVID-19-pneumonia is much rarer. Thus, adopting a training dataset with a balanced distribution among these three categories may further improve the model’s performance. It is also noted that COVID-19 has been proven to be highly mutational, and existing datasets cannot fully account for the evolving variants. This will increase the likelihood for a future COVID-19 patient to receive a false negative diagnosis, which makes multi-round training imperative to ensure the accuracy of AI-assisted COVID-19 diagnostic systems.

4.3. Re-Training Effects on Model Performance and Result Interpretability

In this study, multi-round training not only improved the model performance on test images with diminishing similarity, but also demonstrated the benefits of evaluating the learnability and performance robustness among different models. It is cautioned that retraining with extended data did not always improve model performance, and relevant images with high quality are needed, as shown in Table 3, Table 4, Table 5 and Table 6. The progression in learning behaviors could be observed through the changes in extracted activation features, ROC curves, or heatmaps across different stages. The observation that the ROC curves of ResNet-50 were persistently more skewed to the upper-left corner was closely correlated with its superiority over the other three models at multiple stages (Figure 5 and Table 2, Table 3, Table 4, Table 5 and Table 6). Conversely, VGG-19 consistently presented a lower ROC curve in every round and concurrently exhibited a slower rate of improvement (Figure 5 and Table 2, Table 3, Table 4, Table 5 and Table 6). By tracking alterations in activation features and heatmaps (Figure 8, Figure 9 and Figure 10), we can gain deeper insights into the network’s perception of the disease (i.e., the predominant features for classification) and even enhance our understanding of the disease through features/heatmaps filtered by the CNN models. It should be emphasized that, at least within the context of this study, while a correlation between features/heatmaps and disease manifestation does exist, it remains tenuous. Explicit application of model-learned features/heatmaps to explain the classification is still impractical.

The weak heatmap-disease correlation can be seen in Figure 8 and Figure 9, while the irregularity in such correlations can be seen in the pneumonia case in Figure 9. A normal X-ray image of the lungs will typically show clear, dark spaces with distinct edges that represent healthy air-filled lung tissue [56,57]. There should be no signs of abnormal densities or infiltrations that are indicative of infection or inflammation, as shown in the upper panels of Figure 8 and Figure 9. COVID-19 X-ray images may show patchy areas of opacity, which indicate fluid buildup in the lungs [58,59]. Additionally, COVID-19 may cause a condition called “ground-glass opacity”, which appears as a hazy, cloud-like pattern. In Figure 8 and Figure 9, patches of hotspots with intermediate to high intensity were also observed. X-ray images of pneumonia also show infiltrations and white patches in the lungs [60], but they may appear in a more patchy or lobar pattern, which is consistent with the large extent of hotpots of occlusion intensity in the lower panels of Figure 8. Yet, similar observations are not found in the pneumonia Grad-CAM heatmaps displayed in the lower panel of Figure 9. This suggests that directly using heatmaps to elucidate classification outcomes might not always be viable.

5. Conclusions

Despite reported high accuracies, many previous CNN-based COVID-19 diagnostic models relied on clinic-specific or limited datasets and were rarely tested with external images. In our study, we trained four CNN models on incrementally expanded datasets and assessed their performance using increasingly dissimilar images. By utilizing X-ray images sourced from multiple centers, we evaluated the models’ capabilities in verification, interpolation, and extrapolation. Compared to traditional single-phase training, our multi-stage approach offered a comprehensive evaluation of the models’ learnability, robustness, and interpretability of classification results.