Integration of Free-text Pathology Reports into Models for Automatic Contouring of Clinical Target Volume in Postoperative Head and Neck Cancer

Mu-Hung Tsai, Hsi-Huei Lu, Kun-Hsing Yu, Yuan-Hua Wu, Yung-Jen Cheng, Tzu-Hui Pao, Wei-Ting Hsueh, Irene Tai-Lin Lee, Jung-Hsien Chiang

Keywords: Head and neck cancer, radiotherapy, pathology, clinical target volume

TL;DR: We improved automatic segmentation of clinical target volume by integrating free-text pathology reports.

Abstract:

Background: Deep-learning-based automatic contouring systems for generating clinical target volumes (CTVs) have been proposed for several cancers; however, no studies have reported their implementation in the setting of adjuvant radiotherapy for postoperative head and neck cancer. Additionally, standard image-only deep learning models cannot utilize additional data, such as pathology reports. Therefore, we aimed to investigate model implementation and integration of pathology reports in this setting.

Methods: We extracted the data of 153 patients with head and neck cancer who underwent adjuvant radiotherapy after definitive resection. After evaluating the performance of standard image-only segmentation models, we designed a novel architecture, TransU-Pathology-Net (TransUPNet), to process information from pathology reports alongside image features and evaluated its performance using the Dice similarity coefficient (DSC) and 95th percentile of the Hausdorff distance (HD95). Comparisons were made against the best-performing standard image-only baseline model.

Results: The TransUNet model produced the best performance among the standard image-only models and was selected as the baseline model for comparison. Incorporating pathology data via the TransUPNet architecture improved model performance (DSC: 0.71±0.02, HD95: 33.6±11.6 mm). The importance of pathology input was demonstrated experimentally, showing that perturbations in pathology report data (truncation, laterality switch, and mismatch) resulted in decreased performance.

Conclusions: We developed a deep learning-based system for CTV generation for adjuvant radiotherapy of head and neck cancers. It has the unique ability to integrate free-text information from pathology reports, which improves model performance and more closely mirrors the physician workflow of integrating multiple information sources to generate CTVs.

Primary Subject Area: Integration of Imaging and Clinical Data

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 29