Performance of deep learning-based segmentation of soft tissue sarcoma by MRI sequence, tumor type and location

OBJECTIVE: Soft tissue sarcomas (STS) are a rare and heterogeneous group of tumors that pose a significant challenge for surgical planning. This study evaluated the performance of a deep learning model for automated STS segmentation on preoperative MRI, focusing on how different MRI sequences, anatomical locations, and histological subtypes influence model accuracy. MATERIALS AND METHODS: We retrospectively analyzed 299 patients with biopsy-proven leiomyosarcoma (n = 67), myxofibrosarcoma (n = 55), myxoid liposarcoma (n = 60), undifferentiated pleomorphic sarcoma (n = 70), and intramuscular myxoma (n = 47) with pre-treatment MRI (2004-2022). Tumors were manually segmented on fat-suppressed, contrast-enhanced T1-weighted, and fat-suppressed T2-weighted sequences. Separate 3D nnU-Net models were trained for each MRI sequence and combination. Performance was evaluated using Dice, F2 score, average symmetric surface distance (ASSD), and 95th percentile Hausdorff distance (HD95). RESULTS: Overall, single-sequence T1 models outperformed multi-modal and multi-plane approaches. The T1 axial model achieved the best volumetric accuracy (median F2 score 0.91, Dice 0.89), while the T1 sagittal model provided superior boundary delineation (ASSD 2.1 mm, HD95 4.3 mm). Performance varied by anatomical site: T1 sequences were optimal for tumors in the extremities, whereas T2 sequences were more effective for abdominal and pelvic lesions. Accuracy also varied by histology, with the highest performance for myxofibrosarcoma and myxoid liposarcoma (F2 up to 0.94) and the lowest for leiomyosarcoma (F2 up to 0.88). Unexpectedly, multi-modal fusion of T1 and T2 images did not improve results and often degraded boundary accuracy. CONCLUSION: Deep learning models achieve high accuracy for STS segmentation. However, their performance is critically dependent on tumor location and histology.