Source-linked AI summary
Weakly-supervised localization of diabetic retinopathy lesions in retinal fundus images
Waleed M. Gondal, Jan M. Köhler, René Grzeszick, Gernot A. Fink, Michael Hirsch
TL;DR
Medical experts may distrust CNN predictions because their nonlinear decision process is difficult to interpret, especially when DR evidence is scattered across lesions. The paper adapts a CNN with class activation maps for weakly supervised localization using image-level labels. On DiaretDB1, it reports accurate DR classification and lesion localization with few false positives, while maintaining high sensitivity.
Problem
CNN predictions are difficult for medical experts to interpret, and retinal DR decisions may depend on lesion areas scattered across the image.
Method
The paper modifies a CNN to generate class activation maps for weakly supervised localization using image-level DR labels.
Results
The method achieves 93.6% sensitivity, 97.6% specificity, and 95.4% ROC AUC for binary DR classification on DiaretDB1, with comparable or better lesion-detection sensitivity than supervised methods.
Takeaways & Limitations
The approach highlights DR-indicative retinal regions to assist medical diagnosis while maintaining high classification accuracy and good localization sensitivity.
Abstract
from arXiv · showhide
Convolutional neural networks (CNNs) show impressive performance for image classification and detection, extending heavily to the medical image domain. Nevertheless, medical experts are sceptical in these predictions as the nonlinear multilayer structure resulting in a classification outcome is not directly graspable. Recently, approaches have been shown which help the user to understand the discriminative regions within an image which are decisive for the CNN to conclude to a certain class. Although these approaches could help to build trust in the CNNs predictions, they are only slightly shown to work with medical image data which often poses a challenge as the decision for a class relies on different lesion areas scattered around the entire image. Using the DiaretDB1 dataset, we show that on retina images different lesion areas fundamental for diabetic retinopathy are detected on an image level with high accuracy, comparable or exceeding supervised methods. On lesion level, we achieve few false positives with high sensitivity, though, the network is solely trained on image-level labels which do not include information about existing lesions. Classifying between diseased and healthy images, we achieve an AUC of 0.954 on the DiaretDB1.
1. INTRODUCTION
Diabetic retinopathy screening needs automated support because qualified personnel are scarce, but deep neural networks provide limited insight into their predictions. The paper therefore localizes image regions responsible for DR classification and shows that these regions map to lesions without explicit lesion-level training.
- Diabetic retinopathy screening is important because it can enable early treatment, while a lack of qualified personnel limits retinal-image assessment.
- Deep neural networks achieve good DR performance, but their classification outputs are difficult for medical experts to interpret.
- Existing lesion detectors identify specific lesion categories but cannot directly connect those detections to a deep network’s prediction.
- The proposed method localizes image areas responsible for CNN-based DR conclusions and shows that these areas map to lesion regions without explicit lesion training.
2. RELATED WORK
Weakly supervised localization aims to identify class-specific regions influencing CNN predictions rather than merely visually distinctive areas. Retinal localization is difficult because DR lesions are small and scattered, motivating image-level supervision approaches.
- Weakly supervised object localization highlights class-specific discriminative regions that influence predictions, whereas saliency maps may identify regions unrelated to classification.
- Prior CNN visualization methods include internal-feature visualization, back-projection, masking, and activation-based localization.
- Retinal DR localization is challenging because lesions such as red dots, microaneurysms, and hemorrhages are scattered across the image.
- The cited retina-specific weakly supervised approach uses image-level labels and an ensemble of CNNs, with each CNN specialized for a lesion type.
3. METHOD
The method modifies a DR classification CNN to generate class activation maps through global average pooling, preserving spatial information for weakly supervised localization. It increases feature-map resolution and thresholds heatmaps into lesion-region proposals for evaluation.
- Localization with Class Activation Maps: The CNN removes dense layers and adds global average pooling so spatial feature maps can produce class activation maps while retaining classification scores.
- Localization with Class Activation Maps: Class activation maps combine feature maps using class-specific weights, then upsample the result to highlight discriminative regions in the input image.
- Localization with Class Activation Maps: Feature-map resolution is increased to 32 × 32 pixels by removing strides and adding a 3 × 3 stride-one convolutional layer with 1024 kernels.
- Localization with Class Activation Maps: Batch normalization is introduced to support faster convergence and higher learning rates after spatial-resolution changes slow training.
- Localization with Class Activation Maps: The architecture generates heatmaps for qualitative inspection and converts normalized high-intensity regions into predicted lesion proposals using binary segmentation.
4. EXPERIMENTS
Experiments evaluate image- and lesion-level performance using Kaggle for training and DiaretDB1 for lesion-localization validation. The weakly supervised model achieves strong classification and lesion detection, with comparable sensitivity and fewer false positives than other methods.
- Datasets and Evaluation: The study trained on 88,702 Kaggle fundus images and evaluated lesion localization on 89 expert-labeled DiaretDB1 images.DiaretDB1 contains annotations for four diabetic-retinopathy lesion types, retaining pixels with labeling confidence above 75%.
- Evaluation Criteria: Lesion-level true positives were evaluated using either at least 50% overlap or at least one overlapping pixel with confidence of 0.75 or higher.The authors report similar performance under both criteria.
- Performance at Image Level: 93.6% sensitivity, 97.6% specificity, and 95.4% ROC AUC were achieved for binary RDR classification on DiaretDB1.Specificity is reported over all lesion types because image-level false positives cannot be assigned to a specific lesion type.
- Performance at Image Level: The weakly supervised model performed comparably to or better than fully supervised methods for image-level lesion detection.The comparison concerns methods trained specifically to detect one or two lesion types; sensitivities are reported in Table 2.
- Performance at Lesion Level: The model produced fewer false positives than other methods while achieving comparable sensitivity, although red small dots were difficult to detect.The authors attribute this weakness to information compression in CNNs and coarse heatmap resolution.
5. CONCLUSION
The paper presents a deep learning approach that highlights retinal regions indicative of diabetic retinopathy to assist medical diagnosis. Its modified CNN enables weakly supervised localization while maintaining accurate classification and fast inference.
- 5. CONCLUSION: The approach highlights retinal regions indicative of diabetic retinopathy to assist medical diagnosis.The architecture adapts a top-performing supervised diabetic-retinopathy classification CNN for weakly supervised object localization.