*Proceedings* **Enhancing Retinal Blood Vessel Segmentation through Self-Supervised Pre-Training †**

#### **José Morano 1,2,\*, Álvaro S. Hervella 1,2, Noelia Barreira 1,2, Jorge Novo 1,2 and José Rouco 1,2**


Published: 25 August 2020

**Abstract:** The segmentation of the retinal vasculature is fundamental in the study of many diseases. However, its manual completion is problematic, which motivates the research on automatic methods. Nowadays, these methods usually employ Fully Convolutional Networks (FCNs), whose success is highly conditioned by the network architecture and the availability of many annotated data, something infrequent in medicine. In this work, we present a novel application of self-supervised multimodal pre-training to enhance the retinal vasculature segmentation. The experiments with diverse FCN architectures demonstrate that, independently of the architecture, this pre-training allows one to overcome annotated data scarcity and leads to significantly better results with less training on the target task.

**Keywords:** self-supervised learning; transfer learning; multimodal; retinal vasculature segmentation

#### **1. Introduction**

Retinal vasculature segmentation represents a key step in the analysis of multiple common diseases like glaucoma and diabetes. However, its manual completion is arduous and partly subjective, so automatic methods have emerged as an advantageous alternative. State-of-the-art vasculature segmentation is based on Fully Convolutional Networks (FCNs). Nonetheless, using FCNs requires addressing two major difficulties: (1) Determining the network architecture and (2) gathering a large amount of annotated training data. The first issue can be partly overcome by reviewing similar problems. Annotated data, however, are usually scarce in medical imaging, as they require experts to be involved in a tedious process. This motivates the proposal of self-supervised multimodal pre-training (SSMP) to learn the relevant patterns from unlabeled data and reduce the required amount of annotated data [1–3]. Specifically, the proposed SSMP consists of training an FCN to predict fluorescein angiographies (a grayscale modality that enhances the vasculature) from retinographies.

In this work, we present a novel application of SSMP to enhance vasculature segmentation in a transfer learning setting, performing a comparative analysis of several FCN architectures.

#### **2. Methodology**

The main objective of this work is the segmentation of the retinal vasculature using FCNs. To enhance the results, we propose a transfer learning setting that consists of using SSMP followed by a fine-tuning in the segmentation task [4]. To appraise our proposal, we evaluated the results of the same networks using the SSMP or training from scratch and with different training set sizes (1, 5, 10, and 15). In all of the cases, we used the following FCN architectures: U-Net [5], FC-DenseNet [6], and ENet [7,8].

In order to perform the SSMP, we aligned the 59 retinography–angiography pairs of the publicly available Isfahan MISP dataset [9] using the method proposed in [10]. Then, inspired by [1,2], we used SSIM function to compute the reconstruction loss between the network output and its ground truth.

To train the networks for the vasculature segmentation task, we employed the DRIVE dataset [11], which consists of 40 retinographies and their corresponding vasculature segmentation masks. As the loss, we used Binary Cross-Entropy. For testing, we included the 20 annotated images of the STARE dataset [12].

The networks were trained using the Adam optimization algorithm with learning rate decay and data augmentation through affine transformations and color and intensity variations.

#### **3. Results and Conclusions**

Table 1 shows the best AUC-ROC and AUC-PR values of the different networks trained from scratch (FS) and using SSMP for the STARE dataset. Moreover, in Figure 1 is depicted an example of the segmentation masks predicted by the U-Net trained with 15 images, with and without SSMP. As observed, the use of SSMP has significant benefits in both quantitative and qualitative terms; mainly due to the fact that the vessel continuity is better preserved and the pathological structures are better handled. This improvement, in addition, is achieved with less training in the target task. These results demonstrate that the use of SSMP emerges as a valuable option when annotated data in the target task are scarce.

Regarding the diverse FCN architectures, both qualitative and quantitative results (see Table 1) demonstrated that the U-Net provided the best performance.

**Table 1.** Best AUC-ROC and AUC-PR values of the different networks trained from scratch (FS) and using self-supervised multimodal pretraining (SSMP) for the STARE dataset.


**Figure 1.** Example of the predicted vasculature mask. From left to right: Original STARE retinography, vasculature segmentation ground truth, vasculature segmentation mask predicted by U-Net trained with 15 images without SSMP, vasculature mask predicted by the same network with SSMP.

**Author Contributions:** Conceptualization, Á.S.H., J.N., and J.R.; methodology, Á.S.H. and J.M.; software, Á.S.H. and J.M.; validation, Á.S.H. and J.M.; formal analysis, Á.S.H. and J.M.; investigation, J.M.; resources, N.B., J.N., and J.R.; data curation, Á.S.H. and J.M.; writing—original draft preparation, J.M.; writing—review and editing, J.M. and J.N.; visualization, J.M.; supervision, J.N. and J.R.; project administration, J.N. and J.R.; funding acquisition, J.N. and J.R. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work is supported by the Instituto de Salud Carlos III, Government of Spain, and FEDER funds of the European Union through the DTS18/00136 research projects and by the Ministerio de Ciencia, Innovación y Universidades, Government of Spain through the RTI2018-095894-B-I00 research projects. In addition, this work has received financial support from the Consellería de Educación, Universidade e Formación Profesional, Xunta de Galicia, through the ERDF (80%) and Secretaría Xeral de Universidades (20%), CITIC, Centro de Investigación del Sistema Universitario de Galicia, Ref. ED431G 2019/01.

**Conflicts of Interest:** The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results.

#### **References**


c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Proceedings* **Study on Relevant Features in COVID-19 PCR Tests †**

**Plácido L. Vidal 1,2,\* , Joaquim de Moura 1,2 , Lucía Ramos 1,2 , Jorge Novo 1,2 and Marcos Ortega 1,2**


Published: 26 August 2020

**Abstract:** In the year 2020, the world suffered the effects of a global pandemic. COVID-19 is a disease that mainly affects the respiratory system of patients, even causing a disproportionate response of the immune system and further spreading the damage to other vital organs. The main means by which health care services detected this viral disease was through the use of Polymerase Chain Reactions (PCRs). These PCRs allow the detection of known chains of the genetic code of the virus in samples of sputum. In this work, we study PCR signal features that allow to automatize the analysis of hundreds of PCRs. The findings obtained from the study have shown these features to be capable of obtaining successful results in the detection of COVID-19 in PCR samples, with only a small fraction of the information extracted by the clinicians for that purpose.

**Keywords:** polymerase chain reaction; COVID-19; feature analysis

#### **1. Introduction**

SARS-CoV-2 is a strain of coronavirus responsible for the global COVID-19 pandemic of 2020. This virus causes a severe acute respiratory syndrome (SARS-CoV-2) and viral pneumonia that may leave lungs severely and irreparably damaged [1,2].

For the detection of this virus, sputum samples are analyzed. In these tests, using RT-qPCR or Reverse Transcription Quantitative Polymerase Chain Reaction, the RNA chains that are present in the obtained samples are transcribed into DNA. These tests use deoxyribonucleotides with fluorescent markers that, as soon as they successfully couple with the reference sequence, emit light. This way, by measuring the fluorescence emitted during each cycle, we can know if the reference gene has been found. The gene to be detected is called "E", common to the family to which COVID-19 belongs to group 2 coronavirus.

#### **2. Methodology**

To study these signals, we will analyze different features, in the search for one that best separates positive from negative patients. In this case, we have analyzed the mean, standard deviation and the percentiles 25% and 75% of the PCR signals of the patient (as well as the positive and negative reference signals for the PCR batch of that given sample). To study their separability, we will use a kernel density estimation with a gaussian kernel over a trained support vector machine (SVM) with a lineal kernel.

#### **3. Results**

Our dataset is composed by 65 positive and 65 negative patients. These signals have been generated with a LightCycler 480 Real-Time PCR System from Roche, using fluorophores with wavelength of excitation of 465 nm and wavelength of detection of 510 nm over 45 PCR cycles. In Figure 1, we can see the mean relative fluorescence returned by the positive PCRs and the negative PCRs. This figure clearly shows the general condition that separates a positive sample from a negative one: the excitation of the signal that is obtained after a certain number of PCR cycles. However, this slope and cycle threshold is highly dependent on the batch (thus the need for the reference signals).

**Figure 1.** Mean PCR fluorescence signals per cycle of all positive and negative patients of the dataset.

Nonetheless, the best results were obtained by using only the standard deviation of the main signal, with an F1 score of 0.94. As shown in Figure 2, we can satisfactorily classify the majority of the patients without the necessity of the reference signals of the PCRs and with a minimal overlap between classes.

**Figure 2.** Resulting density estimation for the trained classifier using a kernel bandwith of 0.1.

As future work, we plan a more in-depth study with a larger dataset. This will allow us to generate more complex features, as well as using modern machine learning and statistical strategies to develop a robust system that can, effectively, further speed up the process of diagnosing this disease.

**Author Contributions:** All the authors contributed equally to this work and have read and agreed to the published version of the manuscript.

**Funding:** Instituto de Salud Carlos III, Government of Spain, DTS18/00136 research project; Ministerio de Ciencia, Innovación y Universidades, Government of Spain, RTI2018-095894-B-I00 research project, Ayudas para la formación de profesorado universitario (FPU), grant ref. FPU18/02271; CITIC, Centro de Investigación de Galicia ref. ED431G 2019/01, receives financial support from Consellería de Educación, Universidade e Formación Profesional, Xunta de Galicia, through the ERDF (80%) and Secretaría Xeral de Universidades (20%).

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

#### **References**


c 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
