Matrix Factorization and Deep Learning for Cancer Detection from Histology Images

Histo-pathological images (HI) are often used for the detection of certain tumors. However, analysis of such images is difficult even for experienced pathologists, due to the variability within normal images and also between images showing different types of tumors. Another challenge is that detecting the tumor in histological images and classifying it into different classes consumes time and resources. One of the ways to accelerate processing and analysis is by using automatic and learning-based algorithms. The last few years have led to a particularly intensive development of approaches based on deep learning, in particular convolutional neural networks (CNN), for the classification of histological images and the detection of tumors in such images [1, 2, 3, 4]. However, in the presence of more than two classes, the design of CNN architectures is also a challenge. The main reasons can be summarized into two main items: i) the CNN models proposed in the literature for the detection of tumors in histological images are generally complex and difficult to train and require a large amount of data, which is not always easily achievable due to the paucity of annotated medical data. Besides, training CNN models on large data requires higher computation times and advanced hardware accelerations such GPU and TPU; ii) difficulty to understand intuitively the learned features from the input images, which make the interpretation of the result by the expert no intuitive. In this project, we aim to design a comprehensive deep learning framework to assist the clinician in the process of analyzing histological images, including on one hand the transformation and learning feature for better data discrimination, and on the other hand, an easy interpretation of the data outputs to strengthen the correlation between the results obtained experimentally and the doctor's believe on the expected results. More details on the method will be given in the next section.