Abstract BACKGROUND Deep learning (DL) has significantly advanced autosegmentation techniques in radiation treatment (RT) planning. While DL models for organ-at-risk segmentation are increasingly integrated into clinical workflows, accurate tumor delineation remains a considerable challenge across various cancer types, frequently necessitating manual contouring. Although numerous DL models have shown promise on limited, private datasets, their generalizability and clinical applicability require rigorous, objective evaluation. International benchmarking challenges provide a critical platform for such comparative assessment. This study evaluates the performance of our DL-based segmentation models for distinct brain tumor entities through participation in three challenges hosted at the MICCAI Brain Tumor Segmentation Challenge (BraTS) 2024: Adult Glioma Post-treatment (GLI) 1, Pediatric Tumors (PED) 2, and Meningioma Radiotherapy (MenRT). MATERIAL AND METHODS For the GLI task (n=1350 training, n=188 validation), models were trained to segment adult glioma sub-regions (non-enhancing tumor core, surrounding non-enhancing FLAIR hyperintensity, enhancing tissue, resection cavity, tumor core) from multi-parametric MRIs. The PED task (n=261 training, n=91 validation) involved segmentation of pediatric glioma sub-regions (enhancing tissue, non-enhancing tumor core, cystic component, surrounding non-enhancing FLAIR hyperintensity) from multi-parametric MRIs. The MenRT task (n=500 training, n=70 validation) focused on gross tumor volume (GTV) segmentation in meningioma patients from T1-contrast MRIs. All datasets underwent standardized preprocessing, including intensity harmonization, bias-field correction, and volume cropping. We employed established DL architectures, including SegResNet, MedNeXt, nnU-Net, U-Mamba, and Vision Transformers. Top-performing models were containerized and submitted to the challenge platform for evaluation on independent testing sets. RESULTS Our models demonstrated strong performance across all participating challenges. The proposed methodology achieved first place in the PED challenge, second place in the MenRT challenge, and fourth place in the GLI challenge, highlighting its effectiveness and generalizability. Quantitative evaluation on the testing sets yielded overall lesion-wise Dice similarity coefficients of 0.873±0.17 for GLI, 0.910±0.14 for PED, and 0.815±0.20 for MenRT. CONCLUSION These results indicate that the evaluated models exhibit highly competitive performance, ranking among the top submissions in the BraTS 2024 challenges for brain tumor segmentation and demonstrating potential for clinical translation. Future work aims to integrate these models into a research PACS for prospective evaluation and validation using real-world clinical data.
Astaraki et al. (Wed,) studied this question.