Deep learning–based segmentation of non-perfused volume after magnetic resonance guided high-intensity focused ultrasound therapies: a retrospective evaluation on clinical patient data

INTRODUCTION: Non-perfused volume (NPV) is a key imaging biomarker for treatment efficacy in MR -guided high-intensity focused ultrasound (MR-HIFU). In uterine fibroids therapies, the NPV ratio is strongly associated with clinical outcomes and long-term efficacy. We developed and clinically evaluated a deep learning-based automatic NPV segmentation model to facilitate standardized and efficient post-treatment assessment. MATERIALS AND METHODS: Contrast-enhanced T1-weighted MR images from 79 patients treated with MR-HIFU were retrospectively analyzed, with NPVs manually delineated as the reference. A 2D U-Net convolutional neural network was optimized using a Tversky loss function (α = 0.7) to prioritize volumetric robustness. Performance was evaluated using similarity metrics (Dice coefficient [DC], 95th percentile Hausdorff distance [HD95]) and clinically relevant measures, including Bland-Altman analysis of volumetric quantification error, precision, recall, slice-level missing rate, and processing time. RESULTS: = 12), the model achieved a median DC of 0.900 (IQR: 0.841-0.920) and a median HD95 of 0.42 (IQR: 0.32-0.60) mm. Median absolute NPV quantification error was 4.3 mL (IQR: 2.6-8.4 mL), corresponding to a median relative error of 2.3% (IQR: -2.4% - 10.5%). Precision and recall were 0.909 (IQR: 0.843-0.940) and 0.912 (IQR: 0.881-0.931), respectively, with 2.8% of ground truth-positive slices missed. Average processing time per patient (∼384 slices) was 1-2 min. CONCLUSION: The proposed 2D U-Net provides accurate and time-efficient automatic NPV segmentation. Quantitative and clinically oriented evaluation demonstrates reliable volumetric assessment with minimal slice-level omission, supporting integration into routine MR-HIFU treatment workflows.