Multi-model study of fast VMAT segment dose calculation with deep learning

Objective.Deep learning (DL) methods enable photon dose calculation under two main coordinate representations: Beam's Eye View (BEV) and patient coordinates. We evaluate dose calculation accuracy and speed under these coordinate paradigms and with representative DL models within a unified dataset and pipeline, and introduce two lightweight models for fast photon dose calculation.Approach.Planning computed tomography (CT) scans and volumetric modulated arc therapy plans from 24 prostate cancer patients were used. Monte Carlo simulation generated 5940, 540, and 3053 segment doses for training (11 patients), validation (3), and testing (10), respectively. For BEV, we used a combination of convolutional neural network (CNN) and convolutional long short-term memory network (ConvLSTM) called CNN-ConvLSTM, a CNN-Mamba combination (CNN-Mamba), a transformer-based architecture (DoTA), and a cascaded 3D UNet (C3D). These were trained on CT and segment-projection BEV cuboids. For patient coordinates, the DeepDose individual segment dose prediction framework implemented with C3D (DeepDose-C3D) was trained on cropped CT volumes with four physical inputs. Segment and plan dose accuracy were assessed using local gamma passing ratesγPR(2%/3 mm and 1%/3 mm) and dose-volume histogram metrics. Dose calculation times (inference plus pre/post-processing) were measured on three different graphics processing unit (GPUs).Results.All five models achieved mean localγPRvalues⩾91.0% (2%/3 mm) for segment doses and⩾99.0% (1%/3 mm) for plan doses. Mean per-segment dose calculation times were 79, 67, 298, 490, 356 ms for CNN-ConvLSTM, CNN-Mamba, DoTA, C3D, and DeepDose-C3D, respectively. On the latest-generation GPU available, the corresponding per-plan (average 305 segments) dose calculation times were 5.5, 6.2, 33.6, 38.7, 35.4 s.Significance.Both BEV- and patient-coordinate DL methods achieved accurate photon plan dose calculation, with BEV-based approaches showing more robust segment performance. CNN-ConvLSTM and CNN-Mamba retain comparable accuracy at lower computational cost, enabling fast photon dose calculation.