Multi-task lesion segmentation with a lightweight U$^2$-Net to enhance explainability of mobile screening systems for diabetic retinopathy

Marlin Siebert, Philipp Rostalski

Keywords: diabetic retinopathy, deep learning, multi-lesion segmentation, u-net, fundus image, mobile diagnostic

TL;DR: Evaluating the performance of a light-weight U$^2$-Net implementation for multi-lesion segmentation in fundus images for mobile diabetic retinopathy detection.

Abstract:

In addition to the recent development of deep learning-based, automatic detection systems for diabetic retinopathy (DR), efforts are being made to improve the explainability of those systems, which are usually designed as black-box models.
By providing precise segmentation masks for lesions being related to the severity of DR, a good intuition about the reasoning of the diagnosing system can be given.
Additionally to this progress, the development of light-weight, smartphone-based DR detections systems, being enabled by the growing computing power of edge devices, is of increasing interest in the research community.
Currently, however, only very few diagnosing systems are pursuing both: implementing joint lesion segmentation and disease grading as well as using small, efficient architectures allowing for implementation on edge devices.
In this paper, we evaluate the performance of a lightweight network implementation for lesion segmentation and assess its potential to extend mobile DR-grading systems for improved reasoning.
To this end, the performance of a downscaled U$^2$-Net, a recent derivative of the well-known U-Net, is evaluated and compared in single- and multi-task lesion segmentation to further reduce memory cost from saving redundant models.
Experimental results show promising diagnostic accuracy while maintaining a small memory footprint as well as reasonable inference speed and thus indicate a promising first step towards mobile diagnostic being able to provide both precise lesion segmentation and DR-grading.

Paper Type: validation/application paper

Primary Subject Area: Application: Ophthalmology

Secondary Subject Area: Segmentation

Paper Status: original work, not submitted yet

Source Code Url: only initial work / no new methods yet

Data Set Url: https://ieee-dataport.org/open-access/indian-diabetic-retinopathy-image-dataset-idrid

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.