Multi-Cell Multi-Task Convolutional Neural Networks for Diabetic Retinopathy Grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a significant complication of Diabetes Mellitus, necessitating efficient screening methods.
• Automated analysis of high-resolution retinal images is crucial for detecting subtle pathological changes in DR.
• Training deep neural networks on high-resolution images presents computational challenges, including vanishing/exploding gradients.

Purpose of the Study:

• To develop an effective deep learning framework for automated Diabetic Retinopathy screening.
• To address the computational and gradient issues associated with high-resolution retinal image analysis.
• To leverage the gradual progression of DR stages for improved classification accuracy.

Main Methods:

• Proposed a Multi-Cell architecture that progressively increases neural network depth and input image resolution.
• Implemented a Multi-Task learning strategy combining classification and regression to capture relationships between DR stages.
• Developed Multi-Cell Multi-Task Convolutional Neural Networks (M²CNN) as a generalizable framework.

Main Results:

• Achieved a Kappa score of 0.841 on the Kaggle dataset, ranking 4th among state-of-the-art methods.
• The M²CNN approach demonstrated improved classification accuracy compared to standard methods.
• The proposed architecture effectively handles high-resolution retinal images and mitigates training difficulties.

Conclusions:

• The M²CNN framework offers a robust and accurate solution for automated Diabetic Retinopathy screening.
• The Multi-Cell and Multi-Task learning strategies are effective in improving deep learning performance for medical image analysis.
• The M²CNN is a versatile framework adaptable to various deep neural network architectures for enhanced medical diagnostics.