Performance Analysis of a Single-Input Thermal Image Classifier with Patient Information for the Detection of Breast Cancer

Breast cancer is counted among one of the most invasive cancers with a high mortality rate among women. Early detection is essential as this nature of cancer can be life threatening. Mammography is one of the leading diagnosis techniques. However, there has been increasing research into other methods due to the high cost and painful procedure involved with mammography. Thermography is one such method that has garnered attention in recent years due to its lower cost and non-invasive approach. This work focuses on building and optimizing a single input CNN model for the diagnosis of breast cancer using thermal images when working with a limited dataset. The images were resized from 640 \(\,\times \,\) 480 to 640 \(\,\times \,\) 640 using the Pytorch library. The images of 5 different views are passed to a single CNN model whose output would be a two-element tuple indicating the probability that the patient is diagnosed healthy or sick. Stochastic Gradient Descent with a learning rate of 0.001 has been used as the optimizer. The model can classify 98% of the dataset with a sensitivity and specificity of 1.00 and 0.97, respectively. Hence, for the detection of breast cancer using thermal images, the use of multiple views of a breast along with clinical data is a viable option.