An Efficient Confidence Measure-Based Evaluation Metric for Breast Cancer Screening Using Bayesian Neural Networks

摘要

In this paper, we propose a confidence measure-based evaluation metric for breast cancer screening using a modular network architecture; we use a traditional neural network as a feature extractor with transfer learning, followed by a Bayesian neural network. We show that by providing medical practitioners with a tool to tune two hyperparameters of the Bayesian neural network (fraction of sampled number of networks and minimum probability), the framework can be adapted as needed. We argue that instead of a single number like accuracy, a tuple (accuracy, coverage, sampled no. of networks, minimum probability) can be used as an evaluation metric. We provide experimental results on the CBIS-DDSM dataset, showing accuracy-coverage tradeoff trends while tuning the hyperparameters. To make the proposed framework deployable, we provide source code with reproducible results at https://git.io/JvRqE.