Publications

Abstract : Lack of labeled data is a major issue in the case of disease diagnosis using machine learning or deep learning algorithms. In this chapter, the authors address the issue of lack of labeled training data for medical image diagnosis by providing a completely unsupervised deep learning framework, which requires no labeled data for training. The proposed framework uses a fully unsupervised convolutional neural network called PCANet for feature extraction from medical images followed by classification using the K-means clustering algorithm to identify various diseases. Further, in this work, they analyze the efficacy of the proposed framework by applying it to solve the problem of intracranial hemorrhage (ICH) identification in computed tomography (CT) images in a fully unsupervised fashion. They trained the proposed framework with 1750 computed tomography (CT) slices (around 57 CT scans), without making use of any labels. They evaluated the model with 751 CT slices (around 25 CT scans). The performance of the proposed model during the evaluation phase was as follows: an accuracy of 67%, a weighted average precision of 0.80, a weighted average recall of 0.67, and a weighted average F1-score of 0.72. Therefore, this model performed fairly in the task of intracranial hemorrhage identification despite the small size of the data set used. Since the proposed method works in a fully unsupervised fashion, it eliminates the difficult and time-consuming process of manually annotating each CT slice of the scan with the help of a radiologist. As the proposed framework greatly simplifies the data collection process, it could be preferred for any medical image classification task.