2466 - iRAI-DiT - A Dose Informed Transformer with Early Cross Attention Fusion for Radiotherapy Dose Prediction from in vivo Ionizing Radiation Acoustic Signals

Purpose/Objective(s):

Ionizing Radiation Acoustic Imaging (iRAI) is a non-invasive technique that detects acoustic waves generated by beam-induced tissue heating. This technology has shown potential applications in real-time dose monitoring, registered with the patient’s anatomy, for both conventional and FLASH delivery. Results in phantoms show good clinical precision. However, patient translation remains challenging due to variability during acoustic coupling, probe placement, and patient motion. To address these clinical challenges, we propose an end-to-end machine learning framework that encodes complex treatment planning information and mitigates iRAI noise propagation through heterogeneous anatomy. This goal is achieved by integrating CT anatomy with raw iRAI acoustic signals to generate CT-registered dose delivery maps.

Materials/Methods:

We implemented an iRAI-based Dose Informed Transformer (iRAI-DiT), a multimodal deep learning network for in vivo delivered dose prediction. Network inputs include planning CT volumes and iRAI acoustic signals acquired during irradiation, with treatment planning dose calculated in RayStation as the prediction target. The architecture uses a Vision Transformer to encode CT information, and an acoustic encoder to process raw ultrasound signals. A transformer U-Net then uses fused representations to predict spatially consistent dose distributions.

Model optimization comprised a composite L1, Gamma index, and conformity loss. The study included 8 photon-treated liver patients across 24 fractions. Two hundred stochastic augmentations were generated per fraction, resulting in 4800 training samples. Network performance was evaluated using a leave-out subset of representative patients to address limited sample size and compared with Delay-and-Sum (DAS) reconstruction using Gamma index analysis.

Results:

Preliminary leave-one-out evaluation in patients 1 and 7 demonstrated improved spatial dose agreement relative to DAS. Isodose agreement mean surface distance improved by 31% compared with DAS using iRAI data. At a 10% dose threshold, iRAI DiT increased Gamma passing rates by 65% and 57% under 5 mm 5% and 3 mm 3% criteria, respectively, compared with DAS. Network inference over 1000 samples was performed within 160 ms ± 2 ms using 2 H100 GPUs, demonstrating scalable computational efficiency compatible with future real-time dose prediction workflows.

Conclusion: