| Abstract [eng] |
In recent years, deep learning (DL) methods have been increasingly applied to improve the automation and accuracy of organ-at-risk (OAR) segmentation in radiotherapy planning. Among these, convolutional neural networks (CNNs), particularly U-Net and its extensions, have shown significant advantages over manual contouring by offering greater consistency, reduced inter-observer variability, and faster execution in medical image segmentation tasks. This study systematically evaluated the performance of three 3D DL-based segmentation models: U-Net, Residual Encoder U-Net (ResEncU-Net), and SwinUNETR, on three multi-organ CT datasets: AMOS (Abdominal Multi-Organ Segmentation), BTCV (Beyond the Cranial Vault), KC (dataset provided by The Hospital of Lithuanian University of Health Sciences Kauno Klinikos), with the goal of assessing DL segmentation model suitability for clinical use. Quantitative evaluation based on Dice Similarity Coefficient (DSC), Surface DSC (sDSC), and 95th percentile Hausdorff Distance (HD95) revealed that ResEncU-Net delivered significantly higher segmentation accuracy across datasets, achieving a DSC of up to 0.916, sDSC values exceeding 0.88. U-Net demonstrated strong baseline performance, particularly in more homogeneous datasets such as KC, where its DSC (0.913) was close to that of ResEncU-Net (0.916). However, its segmentation accuracy declined slightly in datasets with greater variability. SwinUNETR, despite its Transformer-based architecture, showed the weakest performance, with largest mean HD95 values (up to 45.95 mm) and inconsistent sDSC scores, especially for small or low-contrast structures. These findings contrast with previously published results that reported strong SwinUNETR performance in large-scale studies, suggesting that its effectiveness is highly dependent on pretraining and dataset size – factors rarely feasible in clinical settings. In conclusion, the findings demonstrated that open-source segmentation models can be effectively integrated into clinical radiotherapy workflows. With expert validation by radiation oncologists, these models can significantly reduce manual contouring time, maintain high segmentation quality, and support more efficient and reproducible treatment planning. This highlights the practical potential of open-source DL-based tools for routine clinical use in radiation oncology, especially when it is tailored to institutional data and workflows. |
