An Ensemble Deep Learning Based Approach for Red Lesion Detection in Fundus Images (1706.03008v2)

Abstract: Diabetic retinopathy is one of the leading causes of preventable blindness in the world. Its earliest sign are red lesions, a general term that groups both microaneurysms and hemorrhages. In daily clinical practice, these lesions are manually detected by physicians using fundus photographs. However, this task is tedious and time consuming, and requires an intensive effort due to the small size of the lesions and their lack of contrast. Computer-assisted diagnosis of DR based on red lesion detection is being actively explored due to its improvement effects both in clinicians consistency and accuracy. Several methods for detecting red lesions have been proposed in the literature, most of them based on characterizing lesion candidates using hand crafted features, and classifying them into true or false positive detections. Deep learning based approaches, by contrast, are scarce in this domain due to the high expense of annotating the lesions manually. In this paper we propose a novel method for red lesion detection based on combining both deep learned and domain knowledge. Features learned by a CNN are augmented by incorporating hand crafted features. Such ensemble vector of descriptors is used afterwards to identify true lesion candidates using a Random Forest classifier. We empirically observed that combining both sources of information significantly improve results with respect to using each approach separately. Furthermore, our method reported the highest performance on a per-lesion basis on DIARETDB1 and e-ophtha, and for screening and need for referral on MESSIDOR compared to a second human expert. Results highlight the fact that integrating manually engineered approaches with deep learned features is relevant to improve results when the networks are trained from lesion-level annotated data. An open source implementation of our system is publicly available online.

Ensemble Deep Learning for Red Lesion Detection in Fundus Images

The paper "An Ensemble Deep Learning Based Approach for Red Lesion Detection in Fundus Images" by Ignacio Orlando et al. presents an innovative methodology for detecting red lesions in fundus photographs, which are critical indicators of diabetic retinopathy (DR). Such lesions include microaneurysms (MAs) and hemorrhages (HEs), which pose a significant challenge for manual detection due to their nuanced visual characteristics. The authors advocate for a technique that blends deep learning with hand-crafted feature extraction to enhance detection accuracy.

Summary of Methods

The proposed approach addresses the challenge of lesion detection through the integration of features obtained via a convolutional neural network (CNN) along with traditional hand-crafted features. Initially, a candidate detection phase extracts potential lesion sites using morphological operations on fundus images. This candidate set serves as input for a round of classification driven by a CNN, which is trained on image patches to capture nuanced pixel relationships and structural information pertinent to lesion characteristics.

This model sets itself apart by augmenting CNN-derived features with domain-specific, manually engineered descriptors. These include both intensity-based and shape-based features, encompassing various facets of lesion appearance, such as mean intensity values, area, perimeter, and other morphological attributes. Both feature sets culminate in the training of a Random Forest classifier, enhancing detection of true positive lesions while reducing false positives.

Key Findings

Empirical validation demonstrates that the ensemble method significantly outperforms either deep learning or hand-crafted descriptors when used independently. In tests conducted on the DIARETDB1 and e-ophtha datasets, the ensemble approach recorded higher per-lesion sensitivity and lower false-positive rates. Notably, the model achieved superior performance metrics, particularly AUC values, when evaluated for diabetic screening and need-for-referral detection against the MESSIDOR dataset, surpassing several well-known benchmarks.

The ensemble approach leverages the complementary strengths of deep learned and hand crafted features, evidenced by the consistent statistical improvement over individual methods. Particularly, the combined model addressed small lesion detection effectively, which is critical for early DR diagnosis.

Implications and Future Directions

The paper highlights the potential of integrating multi-faceted information to address medical image analysis challenges. In particular, ensemble models can fill gaps left by existing deep learning techniques, especially in domains limited by data sparsity and labeling costs. Future research could explore scaling this approach to broader datasets and integrating additional clinical data streams to further enhance diagnostic robustness.

Furthermore, considering the evolving landscape of AI in medical diagnostics, this research underpins the significance of blending domain knowledge with modern learning techniques. Such an approach not only enriches the feature space but also enhances interpretability and deployability in clinical settings, paving the way for more comprehensive and accurate automated screening tools.

The open-source release of the method on GitHub facilitates further development and application across diverse DR datasets, encouraging community-wide collaboration towards better understanding and implementation of automated DR detection systems.