Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading (1904.08764v1)

Published 16 Apr 2019 in eess.IV, cs.CV, cs.LG, and stat.ML

Abstract: Diabetes is a globally prevalent disease that can cause visible microvascular complications such as diabetic retinopathy and macular edema in the human eye retina, the images of which are today used for manual disease screening. This labor-intensive task could greatly benefit from automatic detection using deep learning technique. Here we present a deep learning system that identifies referable diabetic retinopathy comparably or better than presented in the previous studies, although we use only a small fraction of images (<1/4) in training but are aided with higher image resolutions. We also provide novel results for five different screening and clinical grading systems for diabetic retinopathy and macular edema classification, including results for accurately classifying images according to clinical five-grade diabetic retinopathy and four-grade diabetic macular edema scales. These results suggest, that a deep learning system could increase the cost-effectiveness of screening while attaining higher than recommended performance, and that the system could be applied in clinical examinations requiring finer grading.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Jaakko Sahlsten (5 papers)
  2. Joel Jaskari (6 papers)
  3. Jyri Kivinen (1 paper)
  4. Lauri Turunen (1 paper)
  5. Esa Jaanio (1 paper)
  6. Kustaa Hietala (2 papers)
  7. Kimmo Kaski (83 papers)
Citations (178)

Summary

Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading

This paper presents a systematic exploration of a deep learning-based approach for analyzing fundus images to grade diabetic retinopathy (DR) and macular edema (ME), two common microvascular complications due to diabetes. The paper leverages the convolutional neural network architecture Inception-v3, fine-tuning it for specific classification tasks using fundus images of varying resolutions. Importantly, this investigation uses significantly fewer images compared to previous studies, yet achieves comparable or superior results due to higher image resolutions and effective preprocessing.

The primary contribution of this research is its demonstration of an AI system's capability to effectively evaluate DR and ME using five distinct grading systems. The models were trained on varying image resolutions, specifically 256 x 256 up to 2095 x 2095 pixels, and evaluated using metrics such as AUC, accuracy, specificity, sensitivity, and quadratic-weighted kappa. The highest image resolution, 2095 x 2095 pixels, consistently resulted in the best performance across binary and multiclass classification tasks.

For binary classification tasks, such as differentiating referable from non-referable diabetic retinopathy (NRDR/RDR) and macular edema (NRDME/RDME), the model achieved notable results with high sensitivity and specificity even at the largest resolution despite using only a fraction of labeled images typically employed in other studies. The performance indicators include an AUC of 0.987 for NRDR/RDR at 2095 x 2095 pixel resolution, showcasing its reliability against established benchmarks.

The paper also explores multiclass classification tasks using scales like PIRC, PIMEC, and QRDR with impressive macro-AUC scores indicating effective discriminative capacity across different disease severity levels. Training using the largest image resolutions yielded macro-AUC scores of up to 0.991, illustrating robust feature extraction from high-resolution fundus images which potentially allow for finer disease grading.

While the use of fewer training images enhances cost-effectiveness and computational efficiency, it raises important considerations about generalization and the necessity of further validation across diverse populations and imaging setups. The paper acknowledges grader biases and varying clinical standards as potential confounding factors in assessments.

The theoretical implications of this research are immense, suggesting advancements towards deploying deep learning systems in routine DR and ME screenings, potentially alleviating clinical workloads and healthcare costs. Additionally, this paper opens avenues for future research into optimizing model architecture and hyperparameters across varied input image resolutions while maintaining high diagnostic accuracy. The exploration of these computational systems might propel developments in automated diagnostic tools for retinal diseases beyond DR and ME.

Ultimately, the systematic approach established in this research underscores the promising role of machine learning methodologies even under resource constraints, highlighting their potential to refine medical imaging practices and enhance the precision of patient care services.