Multi-Angle Lateral-View Breast Tissue Classification in Thermographic Images Using Deep Reinforcement Learning and Statistical Shape Modelling

Abstract

This study is a multi-stage method for accurately segmenting and classifying breast tissue in thermographic images, with a specific focus on lateral views captured at multiple angles (30{degrees}, 45{degrees}, and 90{degrees}). Our algorithm is a hybrid boundary detection framework that combines scanning, tracking, and a deep reinforcement learning (DRL) algorithm to delineate the external breast tissue boundary precisely. The DRL agent corrects inaccurately detected landmark points and reduces the average localization error to 17.4 pixels. A Statistical Shape Model (SSM) captures shape variability and refines landmark alignment. A geometric arc-based method estimates the posterior breast boundary, which facilitates robust breast tissue isolation across various lateral view angles. After that, we pull texture features using GLCM (Grey Level Co-occurrence Matrix) and feed them through an SVM classifier. The proposed pipeline boosted diagnostic performance - the SVM hit 74.1% accuracy on our segmented images. It is about 15% better than raw originals and 21% above cropped images. The results prove the effectiveness of our multi-angle lateral view analysis combined with reinforcement learning and statistical shape modelling in enhancing breast cancer detection from thermographic imaging.