U-Net based Active Learning Framework for Enhancing Cancer Immunotherapy

Abstract

Deep learning algorithms using convolutional neural networks to segment images have achieved considerable success in recent years and have continued to assist to explore the quantitative measurement of cancer cells in the tumor microenvironment. However, detecting cancerous regions in whole-slide images has been challenging as it required substantial annotation and training efforts from clinicians and biologists. This challenge was overcome by implementing active learning which is a notable instructing process that requires students feedback to partake in the learning cycle effectively. This research adopted the active learning concept by utilizing semantic segmentation deep convolutional neural network model called U-Net which would be further integrated with active learning framework to improve the annotation and training procedure in a feedback learning process. This reduced the amount of time and effort required to analyze the whole slide images. As an active learning strategy, a low-confidence sample selection algorithm was used to improve the learning process. This system selected highly uncertain samples iteratively to strategically supply whole slide images to the training process. The performance results of the proposed approach indicated that the U-Net-based active learning framework has promising outcomes in the feedback learning process.