3040 - Enhancing Deep Learning-Based GTV Segmentation in Head and Neck Cancer: The Critical Role of Target-Specific Manual Registration Refinement

12:30pm - 01:45pm ET

Poster Hall - Exhibit Hall A

Screen: 16

POSTER

Presenter(s)

Seiya Koga, BS - Tohoku University Graduate School of Medicine, Sendai, Miyagi

S. Koga¹, N. Kadoya¹, Y. Katsuta¹, K. Arai¹, T. Hoshino¹, Y. Kimura², Y. Mitsuhashi², and K. Jingu³; ¹Department of Radiation Oncology, Tohoku University School of Medicine, Sendai, Japan, ²AiRato. Inc, Sendai, Japan, ³Tohoku University Graduate School of Medicine, Sendai, Miyagi, Japan

Purpose/Objective(s):

Accurate Gross Tumor Volume (GTV) delineation is paramount in head and neck (H&N) radiotherapy. While multi-modality deep learning (DL) (CT and MRI) is promising, standard automated rigid registration often fails to achieve perfect local alignment at the tumor site due to setup variations or anatomical changes. We investigated whether target-specific manual refinement of registration during training data preparation significantly improves DL segmentation performance compared to standard automated intensity-based registration.

Materials/Methods:

Data from 132 H&N cancer patients (CT and T1-weighted MRI) were analyzed. To evaluate the necessity of manual intervention, two registration workflows were compared: (1) Global Automated Registration (Mutual Information-based) and (2) Target-Specific Manual Refinement (manually adjusted to the GTVp by radiation oncologists). U-Net models were trained for three scenarios: CT-only, CT+MRI (Automated), and CT+MRI (Manual). Performance was quantified using Dice Similarity Coefficient (DSC) and surface Dice (sDSC).

Results:

The model trained with target-specific manual refinement achieved a mean DSC of 0.72 and sDSC of 0.61. In contrast, the automated registration model (DSC 0.69) showed performance comparable to or occasionally worse than the CT-only model (DSC 0.70) in cases of significant spatial misalignment. Notably, manual refinement prevented catastrophic failures observed in automated pipelines; in the most discordant case, manual alignment rescued the model performance by increasing DSC and sDSC by 0.23 and 0.30, respectively.

Conclusion:

Simple multi-modality fusion via global automated algorithms is insufficient for training high-precision AI contouring tools in H&N cancer. Target-specific manual refinement of registration is a vital step in training data curation, ensuring spatial consistency that robustly improves segmentation accuracy and prevents significant AI errors.