Uncertainty-Aware Deep Ensemble Model For Targeted Ultrasound-Guided Prostate Biopsy

Reliable prostate cancer detection using ultrasound imaging can significantly improve patient outcomes and survival. One outstanding challenge for developing machine learning models that can accurately detect cancer is labeling the ultrasound data based on histopathology reports, where only a coarse label is provided for an entire biopsy sample. As a result, training a model for pixel level analysis of ultrasound images with high rate of label noise leads to model overfitting the noisy data and poor generalization performance. To address this challenge, we leverage the deep ensemble technique for uncertainty estimation to detect out-of-distribution inputs and data with noisy labels. We use Masksembles idea to overcome the computational cost associated with the ensemble techniques. We further improve the performance of the model by using co-teaching framework with a regularized loss. We demonstrate that the proposed method helps with the overconfident predictions of learning models, and noisy nature of the labels data.