Multi teacher knowledge extraction for prostate cancer recognition in medical intelligent assistance systems

Abstract

Designing intelligent diagnosis of prostate diseases in intelligent medical assistance systems has gradually become a research hotspot. However, rectal ultrasound (TRUS) as the main diagnostic tool for prostate diseases remains a challenging issue. (1) Due to limited prostate TRUS imaging data, it is difficult to train a robust deep learning model. (2) Compared with TRUS images of other tissues and organs, the visual features of whether the prostate contains cancer in ultrasound images are similar, so it is difficult for a single neural network model to accurately learn the feature representation of the disease. To address the above problems, we first establish a high-quality dataset for prostate TRUS imaging, and then design multi teacher knowledge distillation to achieve accurate disease recognition. The experimental results show that, compared with knowledge distillation without a teacher model and a single teacher model, knowledge distillation using multiple teacher models can significantly improve the accuracy of prostate TRUS image cancer prediction. As the number of teacher models increases, the accuracy rate is further