[論文レビュー] An Improved Model for Diabetic Retinopathy Detection by using Transfer Learning and Ensemble Learning

Md. Simul Hasan Talukder, Ajay Kirshno Sarkar+2 more|arXiv (Cornell University)|Jun 3, 2023

Retinal Imaging and Analysis被引用数 25

ひとこと要約

本論文は、pre-trained CNNs（例: DenseNet variants）を用いた転移学習とアンサンブル学習アプローチを提案し、データ拡張と正則化を組み合わせて、糖尿病性網膜症検出における高い精度を達成し、特定の DenseNet ベースのモデルで 100% の精度と関連指標を報告する。

ABSTRACT

Diabetic Retinopathy (DR) is an ocular condition caused by a sustained high level of sugar in the blood, which causes the retinal capillaries to block and bleed, causing retinal tissue damage. It usually results in blindness. Early detection can help in lowering the risk of DR and its severity. The robust and accurate prediction and detection of diabetic retinopathy is a challenging task. This paper develops a machine learning model for detecting Diabetic Retinopathy that is entirely accurate. Pre-trained models such as ResNet50, InceptionV3, Xception, DenseNet121, VGG19, NASNetMobile, MobileNetV2, DensNet169, and DenseNet201 with pooling layer, dense layer, and appropriate dropout layer at the bottom of them were carried out in transfer learning (TL) approach. Data augmentation and regularization was performed to reduce overfitting. Transfer Learning model of DenseNet121, Average and weighted ensemble of DenseNet169 and DenseNet201 TL architectures contribute individually the highest accuracy of 100%, the highest precision, recall, F-1 score of 100%, 100%, and 100%, respectively.

研究の動機と目的

• Motivate and develop an accurate automatic detector for Diabetic Retinopathy to enable early intervention.
• Leverage pre-trained CNN architectures via transfer learning to improve performance on DR detection.
• Investigate ensemble strategies to boost predictive accuracy and robustness.
• Apply data augmentation and regularization to mitigate overfitting in limited medical imaging data.

提案手法

• Evaluate multiple pre-trained CNN backbones (ResNet50, InceptionV3, Xception, DenseNet121, VGG19, NASNetMobile, MobileNetV2, DenseNet169, DenseNet201) within a transfer learning setup.
• Incorporate pooling and dense layers with dropout at the bottom of architectures to adapt for DR detection.
• Apply data augmentation and regularization techniques to reduce overfitting.
• Build ensemble approaches (average and weighted ensembles) from DenseNet169 and DenseNet201 TL architectures.
• Assess performance using standard classification metrics (precision, recall, F1-score, accuracy).

実験結果

リサーチクエスチョン

• RQ1 Can transfer learning with multiple pre-trained CNNs improve diabetic retinopathy detection accuracy compared to a single model?
• RQ2 Do ensemble strategies among DenseNet-based TL models yield better performance for DR detection?
• RQ3 What regularization and data augmentation strategies most effectively mitigate overfitting in DR datasets?
• RQ4 Which combination of TL architectures provides the best trade-off between precision and recall for DR detection?

主な発見

• DenseNet121 with transfer learning achieves high accuracy in DR detection.
• Average and weighted ensembles of DenseNet169 and DenseNet201 TL architectures reach the highest reported metrics.
• Ensembling and proper regularization lead to strong precision, recall, and F1-score improvements.
• Data augmentation and regularization are employed to reduce overfitting in the transfer learning setup.
• The paper reports 100% accuracy, precision, recall, and F1-score for certain TL configurations (DenseNet169/201 ensembles).