Comprehensive Head and Neck Organs at Risk Segmentation Using Stratified Learning and Neural Architecture Search.

Researchers

Journal

Modalities

Models

Abstract

Organs at risk (OAR) segmentation is an essential step in the radiotherapy of head and neck (H&N) cancer. Automated approaches benefit physicians in significantly reducing manual work and improving the annotation quality and consistency. Current standard deep learning methods face challenges when the number of OARs becomes large, e.g., > 40. Physicians often refer to easy OARs when delineating harder ones, e.g., using mandibles to help identify adjacent salivary glands. This study aims to emulate this process and develop a new approach that stratifies 42 head and neck OARs into anchor, mid-level, and small & hard (S&H) levels to ensure high segmentation quality.We curated two datasets of CT scans, each annotated with 42 OAR masks: one from 142 oropharyngeal cancer (OPX) patients and the other from 31 nasopharyngeal cancer (NPC) patients. Our segmentation method is first developed and evaluated using the OPX dataset and then further tested using the NPC dataset. Emulating clinical practice, we first stratify the 42 OARs into 3 levels. Anchor OARs are high in intensity contrast and low in inter- and intra-reader variability. Mid-level OARs are low in contrast, but not inordinately small. S&H OARs are poor in contrast and very small. For each level, a tailored deep learning segmentation network is developed using the automated network architecture search (NAS). NAS allows the network to choose among 2D, 3D, or Pseudo-3D convolutions. by considering three levels of complexities. Anchor OARs are used to infer the mid-level and S&H OARs segmentation.With 4-fold cross-validation on the OPX dataset, our method has achieved an average 75.1% Dice score (DSC) and 1.1mm average surface distance (ASD). It outperforms the previous leading method, UaNet, on mid-level OAR segmentation by 3.4% DSC increase 0.4mm, ASD reduction; and on S&H OAR of 10.1% DSC increase, 1.0mm ASD reduction, respectively. Using the NPC patients as an unseen testing set, our method has achieved an average DSC of 76.3% and 1.3mm ASD, which is consistent as in the OPX dataset. This result demonstrates the robustness and generalizability of our method in patients, even with various cancer types.We introduced a new stratified method for segmenting a large comprehensive set of H&N OARs. Our method integrates multi-stage segmentation and NAS in a synergy for the first time. It was trained using OPX patients and achieved state-of-the-art performance and generalized well to patients of NPC. Our method is a critical step towards an automated, accurate, and dependable OAR segmentation system in various H&N cancers.Copyright © 2021. Published by Elsevier Inc.