Detection of Diabetic Eye Disease from Retinal Images Using a Deep Learning Based CenterNet Model

Tahira Nazir, Marriam Nawaz, Junaid Rashid, Rabbia Mahum, Momina Masood, Awais Mehmood, Farooq Ali, Jungeun Kim, Hyuk-Yoon Kwon, Amir Hussain

TL;DR

This paper introduces a novel deep learning approach for detecting and classifying diabetic retinopathy (DR) and diabetic macular edema (DME) from retinal images. By combining DenseNet-100 for feature extraction with a customized CenterNet model for object detection, the method achieves high accuracy and robust localization of disease lesions, even in challenging low-intensity or noisy images.

What Problem Does This Paper Address?

Diabetic retinopathy (DR) and diabetic macular edema (DME) are serious eye conditions that can lead to vision loss if not detected and treated early. The current standard for diagnosis involves manual analysis of retinal images by ophthalmologists, which is time-consuming, expensive, and requires specialized expertise. This manual process also faces challenges with scalability, especially as the prevalence of diabetes and, consequently, DR/DME, continues to rise globally. While automated systems exist, they often struggle with generalizing across different disease severities, handling variations in image quality (like low intensity or noise), and accurately identifying early-stage lesions, which are crucial for timely intervention. This paper aims to bridge this gap by developing a more robust and efficient automated system for DR and DME detection and classification.

Methods

The core of the proposed methodology is a custom deep learning framework that integrates a powerful feature extractor with an efficient object detection model.

  • Feature Extraction: The researchers leverage DenseNet-100 as the backbone for feature extraction. DenseNet's architecture, with its dense connections between layers, is known for its ability to reuse features effectively and reduce the number of parameters compared to models like ResNet, leading to computational efficiency and better performance in capturing complex transformations.
  • Object Detection: They adapt the CenterNet object detection framework. CenterNet is a one-stage detector that treats object detection as a keypoint estimation problem, directly predicting the center of an object and then regressing to its size, offset, and other properties. This one-stage approach is generally faster than two-stage detectors (like Faster R-CNN) while maintaining competitive accuracy.
  • Customization: The authors create a "Custom CenterNet" by replacing the traditional ResNet backbone of CenterNet with their chosen DenseNet-100. This integration aims to harness DenseNet's superior feature representation capabilities for the specific task of identifying subtle lesions in retinal images.
  • Annotation and Training: The process involves meticulous dataset preparation, including manual annotation of suspected disease regions to define the precise areas of interest. The custom CenterNet model is then trained on these annotated images. The model employs a multi-task loss function that combines losses for heatmap prediction (locating object centers), dimension prediction (bounding box size), and offset prediction (fine-tuning the center location).

Datasets

The study utilized several publicly available datasets for training and evaluation:

  • APTOS-2019: A dataset containing 3662 retinal images categorized into five DR severity levels (negative, mild, moderate, proliferative, and severe).
  • IDRiD (Indian Diabetic Retinopathy Image Dataset): This dataset includes 516 images with ground truth for both DR (5 classes) and DME (3 classes).
  • EYEPACS: A large and challenging dataset used for cross-dataset validation, comprising 88,704 samples across five lesion classes.
  • Diaretdb1: Another dataset used for cross-dataset validation, featuring four classes of lesions.

Key Insights and Ideas

  • DenseNet-100 for Feature Richness: The choice of DenseNet-100 as the feature extractor is a key insight. Its dense connectivity allows for better gradient flow and feature reuse, which is crucial for capturing the fine details of early-stage DR and DME lesions that might be missed by less sophisticated feature extractors.
  • CenterNet's Localization Power: The paper highlights CenterNet's strength in localization. By framing detection as a keypoint estimation problem, CenterNet can effectively pinpoint lesion centers and then refine bounding boxes, making it adept at identifying even small or subtle abnormalities.
  • Robustness to Image Quality: A significant finding is the method's ability to handle low-intensity and noisy images. This is critical for real-world applications where image quality can vary significantly, and it directly addresses a limitation of many existing automated systems.
  • Single-Stage Efficiency: The use of a one-stage detector like CenterNet provides a computational advantage, making the system more efficient for real-time or large-scale screening applications compared to two-stage detectors.
  • Cross-Dataset Generalization: The successful cross-dataset validation demonstrates the model's ability to generalize well to unseen data from different sources, indicating its robustness and potential for broader applicability.

Key Results

The proposed method demonstrated strong performance across multiple metrics and datasets:

  • High Accuracy: Achieved an average accuracy of 97.93% on the APTOS-2019 dataset and 98.10% on the IDRiD dataset.
  • Superior Localization: Reported an mAP of 0.96 and IoU of 0.973 for DR lesions, and 0.965 mAP and 0.975 IoU for DME lesions.
  • Competitive Speed: The proposed method achieved a test time of 0.21 seconds, which is faster than other state-of-the-art approaches like Faster R-CNN (0.25s) and Mask R-CNN (0.23s) while achieving higher mAP.
  • Comparative Performance: Outperformed existing methods in comparative analyses, showing a significant performance gain (e.g., 5.19% on APTOS-2019 and 24% on IDRiD compared to other methods).
  • Robust Classification: The confusion matrix shows high precision and recall for most classes, with the model accurately classifying various lesion types and disease severities.

Limitations and Future Work

  • Limitations:
    • The paper notes that while the model performs well, it shows slightly lower results for the "no disease" class due to textural similarity with the "microaneurysms" class.
  • The performance on small lesions, while improved, still requires further enhancement.
  • Future Work:
    • The authors plan to evaluate their technique on more challenging datasets.
  • They also aim to adapt the framework for the detection of other eye diseases.
  • Further research into improving the detection of very small lesions is implied.

Summary updated

Abstract

Diabetic retinopathy (DR) is an eye disease that alters the blood vessels of a person suffering from diabetes. Diabetic macular edema (DME) occurs when DR affects the macula, which causes fluid accumulation in the macula. Efficient screening systems require experts to manually analyze images to recognize diseases. However, due to the challenging nature of the screening method and lack of trained human resources, devising effective screening-oriented treatment is an expensive task. Automated systems are trying to cope with these challenges; however, these methods do not generalize well to multiple diseases and real-world scenarios. To solve the aforementioned issues, we propose a new method comprising two main steps. The first involves dataset preparation and feature extraction and the other relates to improving a custom deep learning based CenterNet model trained for eye disease classification. Initially, we generate annotations for suspected samples to locate the precise region of interest, while the other part of the proposed solution trains the Center Net model over annotated images. Specifically, we use DenseNet-100 as a feature extraction method on which the one-stage detector, CenterNet, is employed to localize and classify the disease lesions. We evaluated our method over challenging datasets, namely, APTOS-2019 and IDRiD, and attained average accuracy of 97.93% and 98.10%, respectively. We also performed cross-dataset validation with benchmark EYEPACS and Diaretdb1 datasets. Both qualitative and quantitative results demonstrate that our proposed approach outperforms state-of-the-art methods due to more effective localization power of CenterNet, as it can easily recognize small lesions and deal with over-fitted training data. Our proposed framework is proficient in correctly locating and classifying disease lesions. In comparison to existing DR and DME classification approaches, our method can extract representative key points from low-intensity and noisy images and accurately classify them. Hence our approach can play an important role in automated detection and recognition of DR and DME lesions.
BibTeX
@inproceedings{nazir2021detection,
  author = {Tahira Nazir and Marriam Nawaz and Junaid Rashid and Rabbia Mahum and Momina Masood and Awais Mehmood and Farooq Ali and Jungeun Kim and Hyuk-Yoon Kwon and Amir Hussain},
  title = {{Detection of Diabetic Eye Disease from Retinal Images Using a Deep Learning Based CenterNet Model}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=JjKLgd3TJY},
}
Figure