Deep learning for gamma map prediction in breast IMRT verification

Patient-specific quality assurance (PSQA) in radiotherapy commonly relies on global gamma passing rate (GPR) metrics, which summarize delivery agreement as a single scalar value and therefore provide limited information about the spatial location of dose delivery discrepancies. Spatial prediction of gamma maps may offer a more informative and interpretable representation of PSQA outcomes.

To develop and evaluate a multimodal deep learning framework for predicting two-dimensional gamma maps in breast intensity-modulated radiotherapy (IMRT) verification, using portal dose images, multileaf collimator (MLC) modulation maps, and plan complexity metrics.

A retrospective single-center dataset of 4707 portal dosimetry fields from 644 breast cancer patients treated with IMRT was analyzed. A U-Net-based encoder-decoder architecture was implemented with dedicated encoders for each input modality and intermediate feature fusion before gamma map prediction. The gamma map was predicted as a three-class categorical output (background, failing, and passing) using a softmax layer with per-pixel argmax decoding. Four configurations were compared: portal-only, portal plus modulation maps, portal plus complexity metrics, and fully multimodal input. Five-fold patient-level cross-validation was used to prevent data leakage. Performance was assessed using receiver operating characteristic analysis with a 90% GPR threshold, standard classification metrics, GPR agreement within $\pm 3\%$ and $\pm 5\%$ , and a categorical map-level spatial-agreement analysis between predicted and measured gamma maps.

Predictive performance improved with multimodal integration. The portal-only model achieved an AUC of 0.77 (95% confidence interval [CI] 0.71-0.82), whereas the fully multimodal configuration reached 0.94 (95% CI 0.92-0.96); the difference was statistically significant in a paired bootstrap test ( $p<0.001$ ). Balanced accuracy increased from 0.64 to 0.77. For the fully multimodal model, 87.9% and 94.6% of cases showed agreement between measured and predicted GPR within $\pm 3\%$ and $\pm 5\%$ , respectively. The categorical map-level spatial agreement was $87.00\%\pm 2.08\%$ in fields with measured GPR $<90\%$ , indicating that the model reproduced the approximate spatial distribution of failing regions.

Multimodal gamma map prediction improved breast IMRT PSQA performance compared with portal-only modeling and extended assessment beyond scalar GPR estimation. By providing spatially resolved predictions of failing regions, the proposed approach may support more informative and interpretable pretreatment QA.