Breast Cancer Detection from Thermal Images Using Asymmetries in Learned Texture Vectors

Breast cancer is the most widespread form of cancer among women. While mammography is the gold standard for diagnosis, it is often considered troublesome for the patients. Thermal imaging, on the other hand, has the power to detect early-stage breast cancer without producing radiation or necessitating direct contact with the patient. Since the affected side has an increase in temperature, it is possible to compare the left and right breasts to distinguish healthy cases from malignant ones. We propose a method based on the swapping autoencoder to learn a texture representation that is disentangled from structure and used to effectively detect left-right differences in a classification task. We trained and evaluated our model on the public DMR-IR dataset, in a low data regime compared to existing work with swapping autoencoders. Our model obtained 0.962 accuracy, an F1 score of 0.971 and 0.988 AUROC. It thus achieves competitive results with state-of-the-art methods, but needs fewer parameters and has lower complexity. The code is available through https://github.com/VisionICLab/TADA-SAE.