S2304
Physics - Machine learning and AI algorithms
ESTRO 2026
Results: In comparison between algorithms, the PC algorithm incorporates prior knowledge as constraints, resulting in a more interpretable structure. Two variables, sex and age, were identified as direct causes of overall survival.DAGSLAM, run without prior knowledge, produced a fully directed but more complex structure and didn't identify any direct cause of overall survival. In addition, several atypical edges were observed, such as the mean lung dose to sex, which may reflect measured variables acting as proxies for unmeasured factors. Using the expert-derived DAG as reference, both methods showed limited agreement: PC achieved Precision = 0.20 and Recall = 0.12, while DAGSLAM achieved Precision = 0.24 and Recall = 0.11.
survival in lung cancer’ Radiotherapy and Oncology,206 pp. S3771-S3772. DOI: 10.1016/S0167- 8140(25)00857-6 Available at: https://doi.org/10.1016/S0167-8140(25)00857-6 (Accessed: 2025/11/10). Keywords: Causal Discovery, Cardiotoxcity
Digital Poster 3239
Superior multi-artifact reduction using Mamba infused residual 3D-Unet for adaptive Head and Neck radiotherapy Viktor Rogowski 1,2 , Christian Jamtheim Gustafsson 1,3 , André Haraldsson 1,2 , Per Munck af Rosenschold 1,2 1 Radiation Physics, Department of Hematology, Oncology and Radiation Physics, Skåne University Hospital, Lund, Sweden. 2 Medical Radiation Physics, Lund University, Lund, Sweden. 3 Medical Radiation Physics, Department of Clinical Sciences Lund, Lund University, Malmö, Sweden Purpose/Objective: Adaptive radiotherapy (ART) for head and neck (H&N) cancer requires accurate Hounsfield Units (HU) in
kilovoltage-CT (kVCT) to ensure precise dose calculations. However, kVCT-image quality is
frequently degraded by dental metal artifacts and photon starvation in the shoulder region. Deep learning-based synthetic CT (sCT) generation has emerged as a promising solution for artifact reduction and HU correction in kVCTs. While convolutional neural networks perform well in modeling local image features, they fail to capture global spatial dependencies. Vision transformers (ViT) overcome these limitations but are computationally expensive for 3D medical imaging. The recently introduced Mamba architecture [1] enables efficient global modeling with content-aware feature aggregation and reduced computational costs.This study aimed to develop a novel Mamba-enhanced 3D-Unet architecture for high-fidelity sCT generation with superior multi-artifact reduction capabilities in kVCT images. Material/Methods: A cohort of 264 H&N patients was divided into training (n=202), validation (n=49), and test (n=13) sets. The first daily kVCT served as model input, and the corresponding planning CT as the ground truth target. Three architectures were developed and compared: (i) baseline 3D U-Net, (ii) ViT-based model, and (iii) the proposed Mamba-enhanced model. Both the ViT and Mamba architectures incorporated residual connections and Convolutional Block Attention Modules throughout the encoder and decoder paths for hierarchical feature extraction and channel-spatial attention refinement. Importantly, MAMBA blocks
Fig. 1 Learned DAG via DAGSALM algorithm. mean.lung: mean lung dose; Left_lung_mean: left mean lung dose; CC:cranio-caudal ; Tumour_loc: tumour location.
Fig. 2 DAG via the PC algorithm. Conclusion:
In real-world clinical data, incomplete measurement of causal factors makes it difficult to infer fully reliable DAGs using data-driven discovery alone. These results underscore the need for caution when interpreting inferred causal relations, regarding the validity of model assumptions. References: Summers, M., et al. (2025). ‘2118 application of causal inference techniques for accurate estimation of the causal effect of dose to the heart base on overall-
Made with FlippingBook - Share PDF online