Deep Learning-Based Cosmetic Evaluation of Breast Cancer Conservative Treatment

Cosmetic evaluation of breasts is important for cancer care teams to counsel patients, manage their expectations, and make treatment choices for breast-conserving cancer treatments. However, availability of experts to grade breast cosmesis is a challenge, especially in low and medium income countries (LMICs). We present a technique and a tool for cosmetic evaluation of breast-conserving therapies, including surgery and radiation, which has been evaluated in an LMIC setting. We show that it is more accurate on South Asian and African women than the available alternative that was based on older computer vision techniques and was trained and tested on a Caucasian population. Our technique adapts to domain shifts, such as changes in demographic physiological characteristics and co-morbidity based on deep learning. To reduce the data required to train our technique, we utilized the closest domain in shape and color characteristics with large labeled dataset - human faces - for pre-training. Additionally, we used generative AI for augmenting the available data for South Asian and African populations. We thus demonstrate equitable use of AI is possible with creative use of the data and technology.