Zoom-in-Net: Deep Mining Lesions for Diabetic Retinopathy Detection

Abstract: We propose a convolution neural network based algorithm for simultaneously diagnosing diabetic retinopathy and highlighting suspicious regions. Our contributions are two folds: 1) a network termed Zoom-in-Net which mimics the zoom-in process of a clinician to examine the retinal images. Trained with only image-level supervisions, Zoomin-Net can generate attention maps which highlight suspicious regions, and predicts the disease level accurately based on both the whole image and its high resolution suspicious patches. 2) Only four bounding boxes generated from the automatically learned attention maps are enough to cover 80% of the lesions labeled by an experienced ophthalmologist, which shows good localization ability of the attention maps. By clustering features at high response locations on the attention maps, we discover meaningful clusters which contain potential lesions in diabetic retinopathy. Experiments show that our algorithm outperform the state-of-the-art methods on two datasets, EyePACS and Messidor.

• The paper introduces Zoom-in-Net, a CNN that uses attention maps to focus on crucial areas for diabetic retinopathy detection and lesion localization with image-level supervision.
• Zoom-in-Net demonstrates strong performance on the EyePACS and Messidor datasets, achieving competitive kappa scores and AUC values for detecting diabetic retinopathy.
• This research highlights the effectiveness of attention mechanisms for localizing lesions in medical images using weak supervision, with potential applications beyond diabetic retinopathy.

Overview of "Zoom-in-Net: Deep Mining Lesions for Diabetic Retinopathy Detection"

The paper introduces Zoom-in-Net, a convolutional neural network (CNN)-based algorithm specifically designed to enhance the accuracy and interpretability of diabetic retinopathy (DR) detection from retinal images. Despite being inherently innovative, the strength of this work lies in its pragmatic design, leveraging image-level supervision to streamline both classification and localization tasks. The algorithm's efficiency is demonstrated through its superior performance relative to existing state-of-the-art methods, as evidenced by experiments conducted on two prominent datasets, EyePACS and Messidor.

Key Contributions

The authors make two primary contributions to the field of medical image analysis:

1. Zoom-in Process Emulation: Zoom-in-Net mimics the examination procedure of ophthalmologists by generating attention maps that highlight regions of interest within retinal images. This approach allows the algorithm to focus on and assess suspicious high-resolution patches, improving diagnostic accuracy.
2. Attention Map Accuracy: With a mere four bounding boxes derived from automatically learned attention maps, the algorithm accurately covers 80% of the lesions as labeled by experts. This demonstrates the effectiveness of the attention maps in localizing pathologies without needing exhaustive, manually detailed annotations.

Architecture and Methodology

Zoom-in-Net features a tri-network architecture comprising:

• M-Net: This main network is designed for image classification and processes input images through a series of convolutions and batch normalizations. It produces features that serve as input to the Attention Network.
• A-Net (Attention Network): This network generates score maps and attention maps, the latter yielding a spatial focus on critical regions. These maps are derived using only image-level labels, achieved via the spatial softmax operator, facilitating the localization of pertinent features.
• C-Net (Crop-Network): This component refines the accuracy by analyzing the high-resolution areas identified as suspicious by the A-Net. It integrates features from these patches and the global image to make enhanced DR level predictions.

Experimental Results

The empirical evaluation presents clear evidence of Zoom-in-Net's capabilities:

• EyePACS Dataset: Zoom-in-Net achieves a kappa score of 0.849 on the test set, with an ensemble approach further improving the result to 0.854. This performance is on par with leading solutions in the DR detection challenge.
• Messidor Dataset: The method achieves an AUC of 0.957 for referable/non-referable classification and 0.921 for normal/abnormal classification, underscoring its cross-dataset robustness.

Implications and Future Developments

The study substantiates the viability of employing attention mechanisms for lesion localization in medical images. The findings imply a potential reduction in clinician workload and an increase in detection sensitivity through the automated analysis of retinal images. Beyond DR detection, the architecture's adaptability suggests applications across various medical imaging tasks, offering a replicable method for integrating diagnostic imaging with machine learning.

Future research avenues could explore enhancing the resolution of attention maps for finer-grained localization, adapting the network to different retinal conditions, or integrating additional modalities, such as depth information or temporal sequences, to fortify diagnostic insights.

In summary, the paper affirms Zoom-in-Net as a robust, interpretable, and high-performing tool for DR detection, effectively bridging the gap between classification and localization under weak supervision conditions.