Source-linked AI summary
Microaneurysm Detection in Fundus Images Using a Two-step Convolutional Neural Networks
Noushin Eftekheri, Mojtaba Masoudi, Hamidreza Pourreza, Kamaledin Ghiasi Shirazi, Ehsan Saeedi
TL;DR
Microaneurysm detection matters because MAs are an early sign of diabetic retinopathy, yet automatic detection faces image-quality, scale, vessel-proximity, and class-imbalance challenges. The paper uses a two-stage CNN to select informative candidate patches and classify them. It reports sensitivity of about 0.8 at more than 6 false positives per image, with higher sensitivity than comparison methods on the reported datasets.
Problem
Automatic MA detection faces poor image quality, scale limitations, missed lesions near vessels, and imbalanced training data, despite its relevance as an early DR sign.
Source: §1 · pp. 1–2Method
A two-stage CNN selects informative MA-like candidate patches and then classifies them using a final CNN.
Source: §3.2 · p. 4, §3.3 · p. 5, §6 · p. 7Results
Sensitivity reaches about 0.8 at an average of more than 6 false positives per image, and the method is reported about 0.3 higher than other methods.
Source: §6 · p. 8Takeaways & Limitations
MA detection is relevant to clinical decisions such as assessing patient status and treatment, as well as to pigmentary retinopathy detection.
Source: §1 · p. 1
Abstract
from arXiv · showhide
Diabetic Retinopathy (DR) is a prominent cause of blindness in the world. The early treatment of DR can be conducted from detection of microaneurysms (MAs) which appears as reddish spots in retinal images. An automated microaneurysm detection can be a helpful system for ophthalmologists. In this paper, deep learning, in particular convolutional neural network (CNN), is used as a powerful tool to efficiently detect MAs from fundus images. In our method a new technique is used to utilise a two-stage training process which results in an accurate detection, while decreasing computational complexity in comparison with previous works. To validate our proposed method, an experiment is conducted using Keras library to implement our proposed CNN on two standard publicly available datasets. Our results show a promising sensitivity value of about 0.8 at the average number of false positive per image greater than 6 which is a competitive value with the state-of-the-art approaches.
1 Introduction
The introduction motivates microaneurysm detection as an early sign of diabetic retinopathy with relevance to clinical decisions, while noting persistent challenges in automatic detection. It presents a CNN-based method for selecting and classifying MA candidates.
1 Introduction
- MAs are described as the first symptom of DR and as small red circular spots under 125 micrometers in diameter.Source: §1 · p. 1
- MA detection remains relevant to clinical decisions about patient status and treatment, and is also discussed in relation to pigmentary retinopathy detection.Source: §1 · p. 1
- Earlier approaches include morphology-based candidate detection and classification, as well as neural-network, region-growing, and template-matching methods.Related work describes top-hat processing, matched filtering, shade correction, local vessel segmentation, and supervised pixel classification among the approaches.Source: §1 · p. 2
- Poor image quality, fixed Gaussian-kernel scales, missed lesions near vessels, and class imbalance are identified as challenges for automatic MA detection.The imbalance between non-MA and MA samples can lead to improper training and reduced classification accuracy.Source: §1 · p. 2
- The proposed method selects informative MA-like regions and uses a CNN to distinguish MA from non-MA spots, addressing problems in existing algorithms.The authors state that the approach can decrease false-positive rate.Source: §1 · pp. 2–3
2 Deep Learning in Medical Image Analysis
The section presents deep learning, especially CNNs, as a growing approach in medical image analysis. It describes CNN feature processing and the architecture used for MA detection.
2 Deep Learning in Medical Image Analysis
- Deep learning has begun to benefit medical image interpretation, while CNNs learn image abstractions used in computer vision and medical-image applications.Source: §2 · p. 3
2.1 Convolutional Neural Networks (CNNs)
- CNN layers apply local feature detectors across feature maps, progressing from simple image features to higher-order shape and texture, with probabilities or class labels as outputs.The described MA-detection architecture uses three convolutional and pooling stages, followed by three fully connected layers and a softmax output.Source: §2.1 · p. 3, Figure 1 · p. 3
3 The proposed method
The method uses a two-stage CNN to address image-quality, MA-scale, vessel-proximity, and class-imbalance challenges in microaneurysm detection. A basic CNN selects informative samples for a deeper final CNN, whose probability map is post-processed into MA candidates.
3 The proposed method
- The two-stage strategy uses the basic CNN's probability map to select informative normal samples for a final CNN that classifies pixels as MA or non-MA.The authors position the strategy as addressing poor image quality, fixed Gaussian-kernel scale, MAs near blood vessels, and dataset imbalance.Source: §3 · pp. 3–4
3.1 Preprocessing Step
- Retinal background is estimated by median filtering with a 30 × 30-pixel kernel and subtracted; 101 × 101 patches are labeled by their central pixel.Patches are produced for training the basic CNN.Source: §3.1 · p. 4
3.2 Candidate Selection by basic CNN
- The basic CNN trains on equal numbers of MA and non-MA patches per epoch, then outputs an initial probability map for preparing the final CNN's input dataset.The authors note that this balancing leaves some non-MA samples unused and causes high false positives in initial results.Source: §3.2 · p. 4
3.3 Classification by final CNN
- The deeper final CNN selects informative patches using a thresholded map from the basic CNN and outputs a pixel-wise MA probability map for each test image.The final map is noisy and requires post-processing.Source: §3.3 · p. 5
3.4 Post-processing
- A 5-pixel-radius disk kernel smooths the final probability map, and its local maxima provide candidate MA centroids.The smoothing addresses cases where close candidates merge in the noisy map.Source: §3.4 · p. 5
3.5 The architectures of CNNs
- The basic CNN has three convolution/pooling layers and three fully connected layers, while the final CNN has five convolution/pooling layers and two fully connected layers.Dropout with probability 0.25 is used to reduce overfitting, and the networks use maxout activation.Source: §3.5 · p. 5
4 Experimental Results
The authors evaluate MA detection on ROC and E-Ophtha-MA datasets using cross-validation and compare sensitivity with prior systems. They also report FROC and CPM evaluations, with the CPM table described as confirming the method's success.
4 Experimental Results
- The evaluation uses 100 ROC images and 148 E-Ophtha-MA images, with about 28,786 MA and 258,354 non-MA patches.The ROC images are split into 50 training and 50 testing images; patch augmentation uses mirroring and rotation.Source: §4 · pp. 5–6
- A 75%/25% cross-validation split is exchanged across successive rounds, and sensitivity is compared with prior works on both datasets.An FROC curve plots sensitivity against average false-positive detections per image.Source: §4 · p. 6
- The authors report that Table 5's CPM evaluation confirms the proposed method's success.The supplied passage gives no CPM value.Source: §4 · p. 6
5 Discussion
The discussion compares sensitivity and Competition Performance Measure (CPM) across ROC and E-Ophtha-MA datasets. The proposed method leads on CPM for both datasets, while sensitivity varies with the false-positive rate.
- ROC sensitivity for the proposed method rises from 0.04 at 1/8 FP/image to 0.76 at 8 FP/image.Table 3 covers sensitivities at varying FP/image rates in the ROC dataset.Source: §5 · p. 6, Table 3 · p. 7
- Across the compared methods, Latim has the second-highest sensitivity after the proposed method, while Waikato’s sensitivity ranges from 0.06 to 0.33.The discussion also reports OkMedical and Fujita sensitivities ranging from 0.18 to 0.50; Table 4 covers E-Ophtha-MA sensitivities at varying FP/image rates.Source: §5 · p. 7, Table 4 · p. 7
- On ROC, proposed sensitivity reaches 0.8 and is about 0.3 above other methods; on E-Ophtha-MA, it is about 0.2 above BWu’s method.The discussion describes these comparisons using FROC curves.Source: §5 · p. 7
- CPM is highest for the proposed method on both datasets: 0.45 versus 0.43 and 0.42 for Antal and Lazar on ROC, and 0.42 versus 0.35 for BWu on E-Ophtha-MA.Source: §5 · p. 7
6 Conclusion
The conclusion presents a two-stage CNN for automatic microaneurysm detection and reports favorable sensitivity relative to prior methods. It also outlines future plans to combine the CNN stages and test other medical applications involving imbalanced data.
- The two-stage CNN selects candidates from a balanced dataset and informative image regions, reducing computational complexity.The conclusion describes the method as addressing imbalanced data and inaccurate microaneurysm detection.Source: §6 · p. 7
- Sensitivity is about 0.8 at more than 6 false positives per image, and the method is reported to be about 0.3 higher than other methods.The conclusion also says it can decrease false-positive rate compared with previous methods.Source: §6 · p. 8
- Future work plans to combine the base and final CNNs during training and apply the method to other medical applications involving imbalanced data.Source: §6 · p. 8
7 Compliance with Ethical Standards
The authors report no conflicts of interest and state that the study involved no human or animal subjects. They used the publicly available ROC and E-Ophtha-MA databases.