Publications

Abstract : Training deep learning models to identify diseases like intracranial haemorrhage (ICH) requires a huge labelled dataset. Lack of a huge labelled data set is a common issue for diseases like ICH. To overcome this issue, in this work, we propose a completely unsupervised deep learning framework for the identification of ICH in computed tomography (CT) images. Our proposed method employs unsupervised (principal component analysis) PCA-Net to extract features from CT images. Further, we trained a K-means classifier using the extracted features from PCA-Net to identify ICH in a completely unsupervised fashion, without making use of any class labels. We also trained a supervised linear support vector machine (SVM) classifier using the extracted features from PCA-Net for a comparative study against the K-means algorithm. We trained both PCA-Net with K-means (fully unsupervised) and PCA-Net with linear SVM models using 1750 CT slices. We tested the models with 751 CT slices (88 slices with ICH and 663 slices without ICH). During testing, our proposed method achieved an accuracy of 0.67, a weighted average precision of 0.80, a weighted average recall of 0.67, and a weighted average F1-score of 0.72. Considering the small size of our training set, our proposed unsupervised framework (PCA-Net + K-means) did a fair job identifying ICH in CT images. Thus, our proposed model can act as a completely unsupervised framework for ICH identification in CT images. Thereby, solving the problem of lack of a huge labelled data set for ICH identification using deep learning models.