2238 - A Fully Automated Artificial Intelligence System for Refining Gross Tumor Volume Delineation in Esophageal Cancer Radiotherapy

02:30pm - 04:00pm PT

Hall F

Screen: 14

POSTER

Presenter(s)

Hongfei Sun, PhD - Xijing Hospital, Xian 710032, Shaanxi

H. Sun¹, Z. An², L. Jia³, D. Gao⁴, and L. N. Zhao²; ¹Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, xi'an, Shaanxi, China, ²Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi'an, Shaanxi, China, ³Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, ⁴Shanghai United Imaging Healthcare Co., Ltd, Shanghai, China

Purpose/Objective(s):

Accurate gross tumor volume (GTV) delineation remains critical yet challenging in esophageal cancer radiotherapy, particularly given the substantial inter-observer variability among physicians with different experience levels. This study proposes a dual-task artificial intelligence (AI) system that integrates automated quality evaluation and adaptive modification of GTV delineations.

Materials/Methods:

We retrospectively analyzed 759 esophageal cancer patients (2015-2023) with paired planning CT scans and GTV contours from both senior (=10 years' experience, ground truth) and junior physicians (=2 years' experience). The dataset was partitioned into training (n=615) and testing (n=144) cohorts. The system comprises two modules: 1) Delineation Quality Assessment: Radiomic features from junior physicians' GTVs were processed through LASSO-SVM pipelines to classify contours as requiring major (DSC < 0.6) or minor modifications (DSC = 0.6). The value range of DSC is 0-1; 2) Adaptive Contour Refinement: A hybrid architecture was implemented using nnFormer (enhanced transformer-based architecture) for major modifications (processing CT images and initial contours) and nnUNet (simpler architecture and easy training without sacrificing accuracy) for minor modifications. Performance was benchmarked against a senior-only training model using DSC and HD95 metrics.

Results:

The quality assessment module achieved discriminative capability (AUC=0.74). The refinement system significantly improved baseline performance: major modifications subgroup showed increased DSC from 0.653±0.200 to 0.683±0.106 (p<0.05) and reduced HD95 from 25.683±32.116 mm to 19.200±33.391mm, while minor modifications attained more precision (DSC: 0.802±0.075, HD95: 6.051±6.399mm).

Conclusion:

The results show that the dual-task AI system can effectively identify and classify low-quality delineation results, providing reliable feedback for clinicians. Meanwhile, the system is superior in accuracy to the senior-only training model. It reduces the workload of senior physicians, and also guides junior physicians in modifying delineations. The proposed system offers new insights into intelligent delineation for esophageal cancer.