Deep image mining for diabetic retinopathy screening

Abstract

Deep learning is quickly becoming the leading methodology for medical image analysis. Given a large medical archive, where each image is associated with a diagnosis, efficient pathology detectors or classifiers can be trained with virtually no expert knowledge about the target pathologies. However, deep learning algorithms, including the popular ConvNets, are black boxes: little is known about the local patterns analyzed by ConvNets to make a decision at the image level. A solution is proposed in this paper to create heatmaps showing which pixels in images play a role in the image-level predictions. In other words, a ConvNet trained for image-level classification can be used to detect lesions as well. A generalization of the backpropagation method is proposed in order to train ConvNets that produce high-quality heatmaps. The proposed solution is applied to diabetic retinopathy (DR) screening in a dataset of almost 90,000 fundus photographs from the 2015 Kaggle Diabetic Retinopathy competition and a private dataset of almost 110,000 photographs (e-ophtha). For the task of detecting referable DR, very good detection performance was achieved: A=0.954 in Kaggle's dataset and A=0.949 in e-ophtha. Performance was also evaluated at the image level and at the lesion level in the DiaretDB1 dataset, where four types of lesions are manually segmented: microaneurysms, hemorrhages, exudates and cotton-wool spots. For the task of detecting images containing these four lesion types, the proposed detector, which was trained to detect referable DR, outperforms recent algorithms trained to detect those lesions specifically, with pixel-level supervision. At the lesion level, the proposed detector outperforms heatmap generation algorithms for ConvNets. This detector is part of the Messidor® system for mobile eye pathology screening. Because it does not rely on expert knowledge or manual segmentation for detecting relevant patterns, the proposed solution is a promising image mining tool, which has the potential to discover new biomarkers in images.

Automated retinal lesion detection via image saliency analysis

“…Figure demonstrates the ability our method to detect HE on two randomly selected images from DiaretDB1 and Messidor datasets. In contrast to the MA detection, the HE detection has received relatively little attention, and in the literature performance has been evaluated only on DiaretDB1. Table reports the sensitivity values achieved by different methods on the DiaretDB1 dataset.…”

“…It can be seen that the proposed method achieves the best performance at both the image and lesion levels with the highest sensitivity values of 0.981 and 0.790, respectively. While both the deep‐learning based approaches focus on the detection of class‐specific discriminative regions, the downsampling operator in their architecture results in loss of location information, and the upsampling operator tends to produce a coarse feature map that renders the fine grained lesion localization impossible.…”

“…It is interesting to note that the evaluation metrics demonstrate that our method has better performances than the recent attention attracting deep learning‐based approaches . It is believed that while the latter focuses on the detection of class‐specific features and high classification accuracy, its architecture essential downsampling and upsampling operators imply that it has inherent difficulty in determining the exact location of the features in the original images and thus poor performance in detecting lesion regions at the pixel level.…”

“…Gondal et al proposed to use a modified convolutional neural network (CNN) in weak supervision with only image‐level labels to identify lesions. Quellec et al generated a heat map as a key input for their deep learning approach — ConvNets. This network can be utilized for DR screening, and both bright and dark lesions are detected at the image level and at the lesion level, respectively.…”

“…The majority of previous approaches to lesion detection are based solely on morphological segmentation or classification . These strategies usually work on a single type of lesions with careful parameter optimization and training data collection, but often fail to work for the detection of other types of lesions without problem‐specific optimization or information.…”