Lesion detection and Grading of Diabetic Retinopathy via Two-stages Deep Convolutional Neural Networks (1705.00771v1)

Abstract: We propose an automatic diabetic retinopathy (DR) analysis algorithm based on two-stages deep convolutional neural networks (DCNN). Compared to existing DCNN-based DR detection methods, the proposed algorithm have the following advantages: (1) Our method can point out the location and type of lesions in the fundus images, as well as giving the severity grades of DR. Moreover, since retina lesions and DR severity appear with different scales in fundus images, the integration of both local and global networks learn more complete and specific features for DR analysis. (2) By introducing imbalanced weighting map, more attentions will be given to lesion patches for DR grading, which significantly improve the performance of the proposed algorithm. In this study, we label 12,206 lesion patches and re-annotate the DR grades of 23,595 fundus images from Kaggle competition dataset. Under the guidance of clinical ophthalmologists, the experimental results show that our local lesion detection net achieve comparable performance with trained human observers, and the proposed imbalanced weighted scheme also be proved to significantly improve the capability of our DCNN-based DR grading algorithm.

An Expert Analysis on Two-Stages Deep Convolutional Neural Networks for Diabetic Retinopathy Detection and Grading

This paper presents a methodological approach to address the persistent challenge of automatic diabetic retinopathy (DR) analysis through a sophisticated mechanism involving two-stages deep convolutional neural networks (DCNN). The proposed method aims to surpass existing technologies by providing automated lesion detection in fundus images and evaluating the severity of the diabetic retinopathy. This dual functionality—local lesion detection coupled with global grading—marks a noteworthy advancement, addressing the nuances associated with DR detection and severity assessment.

Core Methodology and Innovations

The algorithm presents an integration of local and global networks designed to learn multifaceted features across different scales pertinent to DR analysis. Two primary innovations are highlighted as differentiators from previous methodologies:

1. Lesion Detection and Severity Grading: The local network is tasked with identifying and categorizing lesions into microaneurysms, hemorrhages, and exudates. These classifications are crucial for understanding the early markers of non-proliferative diabetic retinopathy (NPDR). Pre-processing techniques such as contrast improvement and circular region extraction enhance lesion detection, making them discernible for computational evaluation.
2. Imbalanced Weighting Scheme: The introduction of a weighted lesion map based on the local network's outputs significantly enriches the grading network's capacity to focus on critical lesion patches. This imbalanced attention mechanism allows for a nuanced analysis of fundus images, enhancing the precision and reliability of DR severity predictions.

Quantitative Results

The empirical evaluation presented robust performance metrics. In lesion detection, the local network achieved notably high recall and precision rates, especially in exudate detection with a precision exceeding 0.83. The global network demonstrated superiority in severity grading when employing the imbalanced weighted scheme, evidenced by a Kappa score increase and improved accuracy relative to traditional end-to-end approaches.

Moreover, the grading network's ability to differentiate between referable DR and normal images—demonstrated by an AUC of 0.959—is particularly compelling. This performance metric is an indication of the algorithm's potential to serve as a reliable tool in clinical settings, capable of high-sensitivity and high-specificity operations that meet and exceed human observer performance.

Implications and Future Work

The implications of this paper are multifaceted, with tangible benefits envisioned for clinical practice and theoretical advancements in medical image analysis. The automation of DR screening could significantly reduce the burden on ophthalmologists and provide timely intervention opportunities for patients, particularly in regions with limited access to specialized healthcare services.

From a theoretical standpoint, the two-stages DCNN model can be expanded upon to include additional lesion types, such as cotton wool spots and venous abnormalities, which were not addressed in this research. Furthermore, the inclusion of diabetic macular edema as part of the DR grading process could present further opportunities for refinement and extension of this work.

Conclusion

In summary, this paper delineates a significant contribution to the ongoing development of sophisticated algorithms tailored for diabetic retinopathy detection and grading. While the proposed method offers compelling advantages, future research can expand its applicability and optimize its implementation through enriched datasets and advanced neural architectures. Nonetheless, the research paves the pathway for enhanced diagnostic precision and practicality, aligning technology with critical healthcare delivery needs.