Trustworthy AI : second opinion needed for faster and safer clinical deployment of AI

The development of fast and reliable tools for medical image segmentation is necessary to meet the ever-growing demand in radiography. The models showing the best performance are based on deep neural networks but necessitate a sufficiently large and diverse annotated dataset. This is often unavailable due to the high level of expertise required to annotate the data and privacy regulations. We propose a multi-method and active learning approach to reduce the amount of annotated data needed for training. It compares the predictions of a U-Net model and a Morphons algorithm following a query-by-committee strategy to specifically label the most informative images to add to the training set. Our approach was tested against random selection for bladder and rectum segmentation. We also propose an intelligent system relying on atlas-based segmentation when the limited size of the dataset hinders the performance of the deep learning model. Thus, we show that atlas-based segmentation is a useful second opinion for deep learning models to reduce the amount of annotated data needed and maintain good performance when working with small datasets.