Automatic Identification of Diabetic Macular Edema Using a Transfer Learning-Based Approach ^†

Abstract

This paper presents a complete system for the automatic identification of pathological Diabetic Macular Edema (DME) cases using Optical Coherence Tomography (OCT) images as source of information. To do so, the system extracts a set of deep features using a transfer learning-based approach from different fully-connected layers and different pre-trained Convolutional Neural Network (CNN) models. Next, the most relevant subset of deep features is identified using representative feature selection methods. Finally, a machine learning strategy is applied to train and test the potential of the identified deep features in the pathological classification process. Satisfactory results were obtained, demonstrating the suitability of the presented system to filter those pathological DME cases, helping the specialist to optimize their diagnostic procedures.

1. Introduction

Diabetic Macular Edema (DME) is one of the most prevalent causes of visual loss and blindness in industrialized countries, representing a concerning public health problem. Optical Coherence Tomography (OCT) is a non-invasive diagnostic technique that provides a high-resolution cross-sectional view of the retina, being commonly used for the diagnosis, monitoring and treatment of the DME disease [1,2]. In this way, a precise and automatic classification of OCT scans between normal or pathological DME cases allows the clinical specialists to make a more accurate diagnosis and treatment of this relevant ocular disease. Figure 1 shows representative examples of OCT scans with and without the presence of DME where we can observe a considerable level of deterioration of the main retinal tissues and the consequent thickening of the retina.

2. Methodology

The presented methodology receives, as input, a set of cross-sectional OCT scans centered in the macular region of the retina. As illustrated in Figure 2, the designed pipeline is composed by 3 main stages. Firstly, the method extracts a set of deep features from the OCT scans using a transfer learning-based approach from different fully-connected layers and different pre-trained Convolutional Neural Network (CNN) models [3]. Then, the method identifies the most relevant subset of deep features using different feature selection approaches. Then, a machine learning strategy is applied to generate a classification model. Finally, the method presents, as output, a labelled OCT image with the classification between normal or pathological DME cases.

3. Results and Conclusions

In this paper, we propose a complete methodology for the automatic identification of pathological DME cases using the OCT images as source of information. Satisfactory results were obtained, demonstrating the suitability of the presented system and consequently helping the clinical specialists in their diagnostic procedures, reducing healthcare costs and improving the quality of life of patients with diabetes.