Explainable AI Driven Clinical Thermography for Breast Cancer Detection

Abstract

Explainable AI; Denoising Auto-encoders, Generative Adversarial Networks, Ensemble Learning, Grad-CAM The quest for enhancing breast cancer detection through non-invasive means is of paramount importance, given its global impact on women s health. In pursuit of this goal, the present thesis proposes a comprehensive AI pipeline titled quotExplinable AI for Clinical Thermography,quot aimed at developing an explainable classifier for detecting breast cancer utilizing infrared breast images. This work is anchored on the premise that early detection is critical. While mammography is prevalent, infrared thermography presents a non-contact and non-invasive alternative, which is yet to be fully exploited in clinical settings. The foundation of this thesis is laid by an extensive review of the current literature on IR imaging for breast cancer detection, which has informed the standardization of image acquisition protocols. With this knowledge, four distinct IR breast image datasets conforming to these protocols were amassed, allowing for a multifaceted analysis. The initial contribution of this research is the development of novel denoising auto-encoders using image pyramids to combat the non-local noise inherent in thermographic imaging. By devising a statistical method to ascertain the noise probability distribution, these auto-encoders enhance the image quality. The effectiveness of this approach was validated through experimental results, showing an average Peak Signal-to-Noise Ratio (PSNR) of 50.1 dB and a Noise Reduction Coefficient (NRC) of 0.79. newline