ESTRO 2024 - Abstract Book

S3991

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

this work, we developed deep-learning predictive models to estimate the internal GI gas, a surrogate for anatomical variability from planning CT, from one or more planar radiographs. The models were developed and benchmarked using a public imaging dataset. The best model’s generalisation performance was evaluated on a small independent clinical cohort.

Material/Methods:

CT scans from 353 subjects (aged 5 days to 16 years) from the Paediatric-CT-SEG dataset available on The Cancer Imaging Archive (TCIA) 1 were used in this study. This open dataset consists of paediatric cases from routine indications that exhibit a range of unknown clinical conditions (i.e., not radiotherapy scans). GI gas volume labels were derived from manual GI tract segmentations using a Hounsfield unit (HU) threshold-based method, i.e., isolating the voxels within the GI tract mask with HU<-300 in the original CT image. Volume ranged from 0 to 2266 mL, with the median volume across patients being approximately 186 mL. Orthogonal lateral and frontal digitally reconstructed radiographs (DRRs) were simulated using Plastimatch with default parameters 2 . Four deep-learning architectures were explored 3 : ResNet-18, ResNet-34, ResNet-50, and a 6 layer CNN. Along with single-view input models, dual-view input late feature fusion and ensemble networks were also trained to optimally combine anatomical information from both orthogonal projections (Fig. 1). All models were trained for 200 epochs, with a batch size of 32 and Mean Squared Error loss function. The Adam optimiser was employed, and the learning rate was adjusted using a cosine annealing scheduler. Each model was trained with five different random seeds, using a train/validation split of 254/64 subjects. Model performance was evaluated for each of the five random seeds on a held-out test set of 35 TCIA subjects. An ensemble prediction of the best-performing pipeline was made by averaging the outputs of each of these five models. The generalisability of the ensemble model was further assessed on an independent cohort of five high-risk paediatric abdominal neuroblastoma patients (aged 3 to 19 years) historically treated with external beam radiotherapy. Key metrics such as the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Pearson correlation coefficient (r) were computed.