OICC Press Majlesi Journal of Electrical Engineering 2345-3796 16 4 2022 12 15 Automatic Diagnosis of Breast Cancer in Histopathologic Images Based on Convolutional AutoEncoders and Reinforced Feature Selection 10.30486/mjee.2022.696518 EN Ali Abdulhussain Fadhil College of Medical Technology, Medical Lab Techniques, Al-farahidi University, Iraq Miaad Adnan Anesthesia Techniques Department, Al-Mustaqbal University College, Babylon, Iraq Hamza Radhi College of MLT, University of Ahl Al Bayt, Kerbala, Iraq Mahmood Al-Mualm Department of Medical Laboratories Technology, AL-Nisour University College, Baghdad, Iraq Mahmood Hasen Alubaidy Al-Hadi University College, Baghdad, 10011 Iraq Mohamed Salih Department of pharmacy, Ashur University College, Baghdad, Iraq Sarah Jaafar Saadoon College of Pharmacy, Al-Ayen University, Thi-Qar, Iraq Journal Article 2022 12 15 Breast cancer is one the most ubiquitous types of cancer which affect a considerable number of women around the globe. It is a malignant tumor, whose origin is in the glandular epithelium of the breast and causes serious health-related problems for patients. Although there is no known way of curing this disease, early detection of it can be very fruitful in terms of reducing the negative ramifications. Thus, accurate diagnosis of breast cancer based on automatic approaches is demanded immediately. Computer vision-based techniques in the analysis of medical images, especially histopathological images, have proved to be extremely performant. In this paper, we propose a novel approach for classifying malignant or non-malignant images. Our approach is based on the latent space embeddings learned by convolutional autoencoders. This network takes a histopathological image and learns to reconstruct it and by compressing the input into the latent space, we can obtain a compressed representation of the input. These embeddings are fed to a reinforcement learning-based feature selection module which extracts the best features for distinguishing the normal from the malicious images. We have evaluated our approach on a well-known dataset, named BreakHis, and used the K-Fold Cross Validation technique to obtain more reliable results. The accuracy, achieved by the proposed model, is 96.8% which exhibits great performance.