S1612
Physics - Autosegmentation
ESTRO 2026
Our analysis showed that RectSegmentor with resolution-matching performed for downstream tasks produced more accurate segmentations than nnU-Net on held-out testing, demonstrating better generalization to unseen variations. It also produced a higher percentage of clinically acceptable segmentations compared to the other models. Further analysis is required to establish the utility of this method for radiomic and longitudinal response assessment studies using larger and multi-institutional cohorts. References: [1] Jiang, J., Tyagi, N., Tringale, K., Crane, C. and Veeraraghavan, H., 2022, September. Self-supervised 3D anatomy segmentation using self-distilled masked image transformer (SMIT). In MICCAI (pp. 556-566). Cham: Springer Nature Switzerland.[2] Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J. and Maier-Hein, K.H., 2021. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), pp.203-211.[3] Albert, S., Wichtmann, B.D., Zhao, W., Maurer, A., Hesser, J., Attenberger, U.I., Schad, L.R. and Zoellner, F.G., 2023. Comparison of image normalization methods for multi-site deep learning. Applied Sciences, 13(15), p.8923. Keywords: transformers, deep learning, rectal Enhancing vertebral metastasis segmentation in CT using anatomical context-aware deep learning Mehdi Astaraki 1,2 , Sebastian Pettersson 3,4 , Simone Bendazzoli 5,6 , Paulina Kalinowska 3,7 , Hanna Hargsten 3,7 , Iuliana Toma-Dasu 8,2 , Vitali Grozman 3,7 1 Medical Radiation Physics, Stockholm University, Stockholm, Sweden. 2 Oncology-Pathology, Karolinska Institutet, Stockholm, Sweden. 3 Radiology, Karolinska University Hospital, Stockholm, Sweden. 4 of Molecular Medicine and Surgery, Karolinska Institutet, Stockholm, Sweden. 5 Biomedical Engineering and Health Systems, KTH Royal Institute of Technology, Stockholm, Sweden. 6 Clinical Sciences, Intervention and Engineering, Karolinska Institutet, Stockholm, Sweden. 7 Molecular Medicine and Surgery, Karolinska Institutet, Stockholm, Sweden. 8 Medical Radiation Physics, Stockholm University, Stockholm, Sweden Purpose/Objective: The vertebral spine is a predominant site for metastatic disease from diverse malignancies, Poster Discussion 4812 correlating with significant morbidity and mortality; however, treatments such as SBRT have demonstrably improved survival outcomes. The development of robust, automated detection/segmentation models for vertebral metastases is of great clinical importance.
Nonetheless, the accurate delineation of tiny lesions remains a significant challenge, despite methodological advances in deep learning. This work introduces a novel approach to enhance the segmentation of vertebral metastases intended to support SBRT treatment planning and to facilitate the quantitative monitoring of toxicities associated with radiation therapy. Material/Methods: A retrospective thin-slice CT cohort (N=180) of patients with lytic vertebral metastases was collected (Karolinska Hospital, 2015–2024). An iterative, semi- automated annotation strategy was employed: an initial 35-case batch was manually annotated, radiologist-verified against patient history, and used to train a preliminary model. This model pre-labeled subsequent batches, which were then subjected to radiological reviews and corrections. This iterative procedure was repeated five times until all 180 cases with 446 metastatic lesions were fully labeled. To address the limited dataset size, several established segmentation models [1–3] were pretrained on 1,500 public CTs for vertebra segmentation. The pretrained models were subsequently fine-tuned on the 180-case annotated dataset, incorporating the entire vertebral body and spinal canal as additional labels to provide anatomical context. Model fine-tuning utilized a five- fold cross-validation strategy, and performance was evaluated at the lesion-wise level. Results: Table 1 shows a summary of the segmentation metrics over the validation subsets.
Overlap-based metrics are ill-suited for tiny lesions, as minor segmentation inaccuracies (e.g., a few pixels) can disproportionately skew the results. Nevertheless, the model could detect tiny lesions which were even demanding for expert radiologists. This detection performance was quantified by an excellent overall lesion-wise false positive rate of 1.1% and a false negative rate of 12%. Figure 1 shows an example of the predicted segmentation results for two challenging cases.
Made with FlippingBook - Share PDF online