Related Experiment Video
Updated: Feb 18, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
An ensemble deep learning based approach for red lesion detection in fundus images
José Ignacio Orlando1, Elena Prokofyeva2, Mariana Del Fresno3
1Pladema Institute, UNCPBA, Gral. Pinto 399, Tandil, Argentina; Consejo Nacional de Investigaciones Científicas y Técnicas, CONICET, Argentina.
Computer Methods and Programs in Biomedicine
|November 22, 2017
Summary
This study introduces a new method for detecting diabetic retinopathy (DR) red lesions by combining deep learning and traditional features. This approach improves detection accuracy and aids in early diagnosis of DR.
Area of Science:
- Ophthalmology
- Medical Imaging
- Computer Vision
Background:
- Diabetic retinopathy (DR) is a leading cause of preventable blindness.
- Early signs of DR include microaneurysms (MAs) and hemorrhages (HEs).
- Manual detection of these red lesions in fundus photographs is challenging due to their size and low contrast.
Purpose of the Study:
- To develop an improved computer-assisted diagnosis system for DR red lesion detection.
- To enhance the accuracy and efficiency of DR screening.
- To address the limitations of manual detection and existing automated methods.
Main Methods:
- Proposed a novel red lesion detection method combining deep learned features from Convolutional Neural Networks (CNNs) with handcrafted domain knowledge features.
- Utilized an ensemble vector of descriptors, integrating both feature types.
- Employed a Random Forest classifier to identify true lesion candidates.
Main Results:
- The combined approach significantly outperformed methods using either deep learned or handcrafted features alone.
- Achieved state-of-the-art performance on per-lesion detection in the DIARETDB1 and e-ophtha datasets.
- Demonstrated superior performance for screening and identifying the need for referral on the MESSIDOR dataset compared to a human expert.
Conclusions:
- Integrating manually engineered features with deep learned features is crucial for improving DR red lesion detection when using lesion-level annotated data.
- The developed system offers a promising tool for early and accurate diagnosis of diabetic retinopathy.
- An open-source implementation is available for public use.
