ABSTRACT
This chapter highlights three challenges that are frequently met with when devising a machine learning (ML) solution for medical imaging/computer-aided diagnosis problems: (1) Dependency on the size of the training dataset, (2) class imbalance, and (3) controlling the confidence of the ML model easily. We propose employing state-of-the-art deep learning techniques with some novel ideas that we have employed in the last 3 years to tackle the aforementioned problems. To reduce reliance on training data, we employ transfer learning. To solve for class imbalance, we employ a novel loss function based on focal loss. Finally, we show the application of Bayesian learning to devise a framework to easily control the confidence of an ML framework through easily tunable parameters. The three different solutions are demonstrated through three case studies with extensive experimental results, on Alzheimer’s diagnosis, lesion segmentation, and finally, breast cancer diagnosis.
