Author: Gong Vincent Hao, Daisuke Kawahara, Jokichi Kawazoe, Yuji Murakami, Ikuno Nishibuchi, Peiying Colleen Ruan, Daguang Xu, Dong Yang 👨🔬
Affiliation: Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences, Hiroshima University, Graduate School of Biomedical and Health Sciences, Hiroshima University, NVIDIA 🌍
Purpose:
Accurate tumor segmentation in head and neck cancer is critical for effective treatment planning, but variability in practices across medical facilities poses challenges for standardization. These differences arise from varying protocols and policies. To address this, we partnered with NVIDIA to develop the TAILOR-TS system, which combines semi-supervised learning with a modular design. This system improves segmentation accuracy while adapting flexibly to facility-specific operational protocols, enabling customized solutions.
Methods:
We used three datasets: RADCURE (2,994 patients), SegRap (200 patients), and an institutional dataset (120 patients). Model performance was evaluated through three approaches: (1) training models separately on each dataset, (2) fine-tuning models between datasets, and (3) implementing TAILOR-TS, which incorporates labeled data from the institutional dataset and unlabeled data from RADCURE. Performance was assessed using the Dice Similarity Coefficient (DSC).
Results:
Training models independently (Scenario 1) yielded DSCs of 0.500 (RADCURE), 0.730 (SegRap), and 0.37 (institutional dataset). With TAILOR-TS (Scenario 2), DSCs improved to 0.757 (SegRap) and 0.42 (institutional dataset) using 100 images, and to 0.782 and 0.54, respectively, with 1,000 images. Using the full RADCURE dataset (2,994 images), DSCs reached 0.781 (SegRap) and 0.62 (institutional dataset). The results highlight that datasets with higher baseline DSC (e.g., SegRap) achieve near-optimal performance with fewer images, while datasets with lower baseline DSC benefit from larger data volumes, emphasizing the need for adaptive strategies based on dataset characteristics.
Conclusion:
TAILOR-TS enhances segmentation accuracy with limited labeled data, offering a practical and adaptable tool for head and neck cancer treatment. By addressing inter-facility variability, it delivers customized solutions to improve treatment planning.