Source-linked AI summary
Ensemble of Deep Convolutional Neural Networks for Learning to Detect Retinal Vessels in Fundus Images
Debapriya Maji, Anirban Santara, Pabitra Mitra, Debdoot Sheet
TL;DR
Robust, exhaustive detection of retinal vessels in color fundus images remains a challenge for computer-assisted screening. This paper trains an ensemble of ConvNets on fundus images and averages their pixel-level vessel probabilities. On DRIVE, the reported ROC area is 0.9283.
Problem
Robust and exhaustive retinal-vessel detection in color fundus images remains a challenge for computer-assisted diagnosis and routine screening.
Source: §I · p. 1Method
An ensemble of ConvNets is trained on raw color fundus images to classify vessel versus non-vessel pixels, with their inference probabilities averaged.
Source: §I · pp. 1–2Results
On DRIVE, the proposed method has an area under the ROC curve of 0.9283.
Source: §V · p. 3Takeaways & Limitations
The authors present a heuristics-independent, data-driven approach as an alternative for medical image analysis.
Source: §VI · pp. 3–4
Abstract
from arXiv · showhide
Vision impairment due to pathological damage of the retina can largely be prevented through periodic screening using fundus color imaging. However the challenge with large scale screening is the inability to exhaustively detect fine blood vessels crucial to disease diagnosis. In this work we present a computational imaging framework using deep and ensemble learning for reliable detection of blood vessels in fundus color images. An ensemble of deep convolutional neural networks is trained to segment vessel and non-vessel areas of a color fundus image. During inference, the responses of the individual ConvNets of the ensemble are averaged to form the final segmentation. In experimental evaluation with the DRIVE database, we achieve the objective of vessel detection with maximum average accuracy of 94.7\% and area under ROC curve of 0.9283.
I. INTRODUCTION
Fundus imaging supports early screening for retinal disease, but robust and exhaustive vessel detection remains a challenge for computer-assisted diagnosis. The paper addresses this with an ensemble of ConvNets trained on raw color fundus images to distinguish vessel from non-vessel pixels.
I. INTRODUCTION
- Robust and exhaustive detection of retinal vessels in color fundus images is an important challenge for computer-assisted diagnosis and routine screening.Clinical lesion reports locate abnormalities relative to vessels or the optic disc.Source: §I · p. 1
- Existing methods rely largely on handcrafted features or heuristic assumptions, while recent data-driven deep-learning models were reported as weaker than state-of-the-art conventional methods.The stated challenge is an end-to-end framework that learns representations from data to identify coarse and fine vessels, at least matching heuristic-based models.Source: §I · p. 1
- The proposed approach trains an ensemble of 12 ConvNets on raw color fundus images to distinguish vessel pixels from non-vessel pixels.Each network is trained independently on randomly selected patches; inference averages their vesselness probabilities into a per-pixel probability.Source: §I · pp. 1–2
- The paper evaluates the approach on the DRIVE dataset and presents its results in the experimental-results section.The introduction also notes that the paper formally defines the problem and describes the approach in separate sections.Source: §I · p. 2
II. THEORETICAL BACKGROUND
The background describes ConvNets as image-processing networks built from feature-transforming layers, and ensemble learning as a way to combine multiple models’ predictions. It introduces these concepts as foundations for the proposed solution.
II. THEORETICAL BACKGROUND
- ConvNets use convolutional, pooling, fully connected, and ReLU layers; convolutional filters transform feature maps, while pooling downsamples them and supports invariance to small translations.Fully connected layers use incoming weights and a bias; ReLU performs pointwise rectification as an activation function.Source: §II · p. 2
- In deeper ConvNet layers, units indirectly interact with larger input areas, forming higher-level abstractions of the input.Source: §II · p. 2
- Ensemble learning combines multiple models’ outputs, for example by averaging, and seeks model diversity to reduce overfitting; it can promote generalization and higher prediction accuracy than individual models.Source: §II · p. 2
III. PROBLEM STATEMENT
The problem is to model the probability of a vessel or non-vascular tissue label at each location in an RGB fundus image. The proposed model class is an ensemble of ConvNets.
III. PROBLEM STATEMENT
- The input is an RGB fundus image, and the pixel labels are vessel and non-vascular tissue.Source: §III · p. 2
- At location x, the method models the probability of tissue label ω using image I and the local neighborhood N(x).Source: §III · p. 2
- The model class H uses parameters learned from training data and hyperparameters tuned using validation data; here, H is an ensemble of ConvNets.Source: §III · p. 2
IV. PROPOSED SOLUTION
The proposed ensemble uses ConvNets that process color fundus patches through convolutional, pooling, and affine layers to produce class probabilities. The layer organization is shared across models, while the penultimate affine-layer width varies.
IV. PROPOSED SOLUTION
- Each ConvNet processes a 3 × 31 × 31 color fundus patch through three convolutional layers, two fully connected layers, and a final softmax.The layer sequence includes ReLU activations, pooling, and dropout before the softmax.Source: §IV · p. 2
- Convolutional layers use 4 × 4 receptive fields, stride 1, and output volumes of 64 × 32 × 32; pooling uses 2 × 2 receptive fields and stride 2.Dropout deactivates a fraction of neurons during learning to discourage co-adaptation and promote generalization.Source: §IV · p. 2
- Figure 2 diagrams the ensemble ConvNets’ layer organization; n denotes the variable number of hidden units in the penultimate affine layer.The caption also defines S as stride and P as zero padding.Source: §IV · p. 2, Figure 2 · p. 3
V. EXPERIMENTAL RESULTS AND DISCUSSION
On DRIVE, the proposed method achieved an area under the ROC curve of 0.9283. Its accuracy was not the highest among compared methods, but it exceeded the earlier deep-learning vessel-representation method.
V. EXPERIMENTAL RESULTS AND DISCUSSION
- The ensemble was learned using DRIVE images 21–40 and tested on DRIVE images 1–20..Source: §V · p. 3
- The method did not have the highest accuracy, but outperformed the earlier deep-learning method for learning vessel representations.The reported kappa score was used to indicate sensitivity to both coarse and fine vessels.Source: §V · p. 3
- Each ConvNet was trained independently on 60,000 randomly chosen 3×31×31 patches, with dropout probability, L2 regularization, and penultimate-layer width sampled across models.The learning rate and annealing rate were fixed at 5e−4 and 0.95, respectively.Source: §V · p. 3
- 0.9283 was the reported area under the ROC curve for the proposed method.Figure 4 presents the ROC curve.Source: §V · p. 3
VI. CONCLUSION
The conclusion presents a ConvNet-ensemble framework for coarse and fine vessel detection, evaluated on DRIVE. It characterizes the approach as data-driven and independent of heuristics, and reports improvement over prior fundus-image work.
VI. CONCLUSION
- The paper presents a ConvNet ensemble for detecting coarse and fine vessels and evaluates it experimentally on DRIVE.The conclusion describes deep learning combined with ensemble learning as an alternative for complex medical data analysis.Source: §VI · pp. 3–4
- Figure 3’s caption identifies coarse-vessel detections in Sample #5 and coarse- and fine-vessel detections in Sample #16 [7].The caption assigns panels a–d to Sample #5 and panels e–h and i–l to Sample #16.Source: Figure 3 · p. 3
- The approach leverages ConvNets’ image recognition and ensemble learning to provide a data-driven, heuristics-independent approach to medical image analysis.The authors describe it as a feasible solution to subjectivity-induced bias in medical image analysis.Source: §VI · pp. 3–4
- The authors characterize this approach as an improvement over their earlier data-driven fundus-image analysis work.They also present deep learning combined with ensemble learning as an alternative for complex medical data analysis.Source: §VI · p. 4