A ResNet is All You Need? Modeling A Strong Baseline for Detecting Referable Diabetic Retinopathy in Fundus Images

TL;DR

This study demonstrates that a standard ResNet-18 model, with careful training and data augmentation, can serve as a strong baseline for diabetic retinopathy detection, achieving competitive performance without complex innovations.

Contribution

The paper shows that a simple ResNet-18, trained with proper preprocessing and augmentation, can match or outperform more complex models for diabetic retinopathy detection.

Findings

ResNet-18 achieved an AUC of 0.955 on a large public dataset.

The model demonstrated good generalization on in-house datasets.

Careful training and data augmentation are key to strong performance.

Abstract

Deep learning is currently the state-of-the-art for automated detection of referable diabetic retinopathy (DR) from color fundus photographs (CFP). While the general interest is put on improving results through methodological innovations, it is not clear how good these approaches perform compared to standard deep classification models trained with the appropriate settings. In this paper we propose to model a strong baseline for this task based on a simple and standard ResNet-18 architecture. To this end, we built on top of prior art by training the model with a standard preprocessing strategy but using images from several public sources and an empirically calibrated data augmentation setting. To evaluate its performance, we covered multiple clinically relevant perspectives, including image and patient level DR screening, discriminating responses by input quality and DR grade, assessing model uncertainties and analyzing its results in a qualitative manner. With no other methodological innovation than a carefully designed training, our ResNet model achieved an AUC = 0.955 (0.953 - 0.956) on a combined test set of 61007 test images from different public datasets, which is in line or even better than what other more complex deep learning models reported in the literature. Similar AUC values were obtained in 480 images from two separate in-house databases specially prepared for this study, which emphasize its generalization ability. This confirms that standard networks can still be strong baselines for this task if properly trained.