S1578
Physics - Autosegmentation
ESTRO 2026
Segmentation accuracy was evaluated by surface Dice Similarity Coefficient with 3 mm threshold, and the 95th percentile Hausdorff Distance. Results: The ITVBPs-DL and the ITVAvgCT-DL approach achieved surface DSC values of 0.69 (0.20) and 0.59 (0.24), respectively, which were significantly different (p<0.01). The lower performance of the model on the average scan may be related to blurring of the tumor image due to breathing motion , which is especially prominent for tumors close to the diaphragm (Figure 1).
Digital Poster 3017 Context-aware brain tumor segmentation: winning solution of MICCAI BraTS lighthouse challenge Mehdi Astaraki 1,2 , Iuliana Toma-Dasu 1,2 1 Medical Radiation Physics, Stockholm University, Stockholm, Sweden. 2 Oncology-Pathology, Karolinska Institutet, Stockholm, Sweden Purpose/Objective: Advances in deep learning (DL) methodology have facilitated the development of robust solutions for automatic tumor segmentation when applied to limited or private datasets. A critical challenge persists, however, in establishing the generalizability and translational utility of these models, which demands rigorous and objective evaluation. In this work, we introduce a novel approach to enhance the segmentation of diverse brain tumor entities by integrating healthy brain structures as anatomical contexts. To evaluate the robustness of this approach objectively, we participated in the MICCAI BraTS 2025 Lighthouse cluster of challenges [1]. Material/Methods: We propose incorporating healthy brain tissues as distinct classes alongside primary tumoral regions within the segmentation labels. The central hypothesis posits that enforcing the model to explicitly differentiate healthy tissues, white matter, gray matter, and CSF, from pathological areas, will enhance overall segmentation accuracy. This multi-class formulation, however, inherently introduces a severe class imbalance. To mitigate this challenge, an adaptive Dice (DSC) loss function was employed. It is derived as an extension of the Cosine similarity function, dynamically modulating its similarity measurement parameters based on the validation accuracy[2]. This proposed pipeline was integrated into the established nnU-Net framework [3].This method was applied to several brain tumours, including glioma, meningioma, and brain metastasis. The aggregated datasets consisted of 5,658 training, 888 validation, and undisclosed testing cases. Results: The proposed method achieved high performance on all the brain tumours in which it was applied and ranked first in the brain metastasis (METs), pre- operative Meningioma (MEN-pre), Meningioma radiotherapy (MEN-rt) BraTS 2025 challenges and secured a high-ranking fourth-place finish in the Adult Glioma (GLI) pre- and post-treatment, as well as pediatric (PED) challenge.Table 1 shows the performance of the evaluated models on the validation sets. Figure 1 shows some examples of the predicted segmentation masks.
The segmentation accuracy for GTVMaxExp-DL and ITVBPs-DL is very similar (Table 1) indicating that the performance of the DL GTV segmentation model that was trained on the max expiration phase is consistent across all respiratory phases.
Conclusion: Deep learning segmentation of the ITV based on CT scans of the different breathing phases results in significant better performance than using the Average CT. The performance of this model is promising, but needs further improvement before it can be used in clinic. References: [1] Ma Y, Mao J, Liu X, et al. Deep learning-based internal gross target volume definition in 4D CT images of lung cancer patients. Med Phys 2023;50:2303–16. https://doi.org/10.1002/mp.16106.[2] Alessia De Biase et al Deep learning aided oropharyngeal cancer segmentation with adaptive thresholding for predicted tumour probability in FDG PET and CT images 2023 Phys. Med. Biol. 68 055013 Keywords: Deep Learning, Lung Cancer , ITV Segmentation
Made with FlippingBook - Share PDF online