Comprehensive and clinically accurate head and neck cancer organs-at-risk delineation on a multi-institutional study

Abstract

Accurate organ-at-risk (OAR) segmentation is critical to reduce radiotherapy complications. Consensus guidelines recommend delineating over 40 OARs in the head-and-neck (H&N). However, prohibitive labor costs cause most institutions to delineate a substantially smaller subset of OARs, neglecting the dose distributions of other OARs. Here, we present an automated and highly effective stratified OAR segmentation (SOARS) system using deep learning that precisely delineates a comprehensive set of 42 H&N OARs. We train SOARS using 176 patients from an internal institution and independently evaluate it on 1327 external patients across six different institutions. It consistently outperforms other state-of-the-art methods by at least 3–5% in Dice score for each institutional evaluation (up to 36% relative distance error reduction). Crucially, multi-user studies demonstrate that 98% of SOARS predictions need only minor or no revisions to achieve clinical acceptance (reducing workloads by 90%). Moreover, segmentation and dosimetric accuracy are within or smaller than the inter-user variation.

vOARiability: Interobserver and intermodality variability analysis in OAR contouring from head and neck CT and MR images

“…27 The reported interobserver variability serves as a baseline performance that any auto-segmentation tool should aim to surpass. An efficient tool should, in practice, outperform the interobserver variability by leveraging the ability of deep learning models [18][19][20][21][22][23][24][25][26] to learn from multiple contouring styles, and generate segmentation masks that are more representative of both observers. Furthermore, the investigation of the intermodality variability provides important information regarding the visibility and distinguishability of OARs as represented by different imaging modalities.…”

“…[6][7][8][9] To reduce observer variability for OARs in the HaN region, several contouring guidelines 4,10 have been published, as well as initiatives have been recently launched to quantify observer variability [10][11][12] and provide quality assurance. [13][14][15][16] On the other hand, automated contouring (i.e., automated segmentation, auto-segmentation) performed by computerassisted algorithms 17 has witnessed a revival with the introduction and integration of artificial intelligence approaches, such as deep learning, [18][19][20][21][22][23][24][25][26] which has outperformed the previously established atlas-based auto-segmentation. 27 As a result, computational challenges were organized to evaluate the quality of auto-segmentation results, 28 and several datasets were made publicly available for benchmarking different auto-segmentation methodologies 20,[28][29][30][31] and evaluating their clinical acceptability.…”

“…In this study, we therefore analyze the interobserver and intermodality variability of manual contouring of up to 31 OARs in the HaN region, performed by observers with different level of experience from CT and MR images of the same patients.Besides providing valuable insights to the levels of both interobserver and intermodality variability from the perspective of manual OAR contouring, the obtained results can be also viewed as a baseline for an objective evaluation of methods for auto-segmentation of OARs in the HaN region, 31 which have been rapidly evolving during the past decade due to the integration of artificial intelligence, and received a considerable boost in performance due to the advances in deep learning. [18][19][20][21][22][23][24][25][26] F I G U R E 1 Example of a CT (top row) and MR (bottom row) image of the same patient from the devised cohort, displayed in mid-axial (left), mid-coronal (middle), and mid-sagittal (right) cross-sections. (For visualization purposes, the original cross-sections were zoomed to the region of interest.)…”

Dilated Adversarial U-Net Network for automatic gross tumor volume segmentation of nasopharyngeal carcinoma

“…In past clinical practices, the delineation of OARs and GTVs in NPC was predominantly conducted by experienced radiation oncologists. However, according to the clinical treatment guideline, each patient has more than 40 OARs and 2 GTVs need to be delineated accurately (Ye et al, 2022;Guo et al, 2020). It requires the radiation oncologists to spend much time performing delineation, increasing the annotator's burden and patient waiting time.…”

Continual Segment: Towards a Single, Unified and Accessible Continual Segmentation Model of 143 Whole-body Organs in CT Scans

“…Multi-organ segmentation has been extensively studied in the medical imaging because of its core importance for many downstream tasks, such as quantitative disease analysis [23], computer-aided diagnosis [44], and cancer radiotherapy planning [25,59]. With the emergence of many dedicately labeled organ datasets [2] and the fast developments in deep learning segmentation techniques [22], deep segmentation networks trained on specific datasets achieve comparable performance with human observers [48,51,59]. However, this setup can have serious limitations in practical deployment for clinical applications.…”