Addressing Label Variability in Deep Learning-Based Segmentation for Radiation Therapy
Accurate segmentation of tumor and surrounding organs-at-risk (OARs) is important for radiotherapy treatment(RT) planning. Manual segmentation by physicians currently used in clinical practice is time consuming and highly depends on the physicians' skill and experience, leading to large inter-and intra-observer variation. Deep learning(DL) algorithms, such as those used for image recognition, has been promising in the development of automated segmentation tools for medical imaging. Contouring in RT has some specific challenges as opposed to semantic segmentation in other spaces. DL segmentation models are trained with large, annotated datasets. The annotations play such an important role since these labels are the supervisory signals teaching the model how to segment. Majority of the times, a very high-quality dataset is unavailable. The datasets available are usually biased, noisy, and sometimes scarcely annotated. One of the primary sources of noise in RT datasets is annotation variations. This dissertation addresses the challenge of label variability and investigates different methods to deal with various kinds of label variability. The thesis studies the impact that label variability has on automation performance as well as patient outcome and devices ways to reduce or to respect this variability. Computed Tomography based segmentation model for intact and post-operative prostate cancer RT planning are proposed and developed for clinical use. Uncertainty in the automated label prediction for CTV and OARs are evaluated in detail. Impact of inter-physician variability on patient outcome for post-operative prostate cancer is investigated and instead of segmenting a general CTV , PSA-Net is proposed that respects style variations across physicians as well as across institutions. For dealing with systematic label variability in some structures, a prior-guided deep difference meta-learner is proposed that can segment a structure in a new labelling style from just a handful of prior segmented patients. A multi-modality IPA segmentation model is proposed to reduce label variability due to expertise differences among physicians in clinical trials. This model can effectively help inexperienced physicians in producing expert segmentations. The contributions of this thesis are expected to facilitate better understanding of label variability in RT and help in avoiding/respecting label variability when developing deep learning models for RT.