In MR-guided high-intensity focused ultrasound (MR-HIFU) treatment for uterine fibroids, the subcutaneous abdominal fat layer is prone to unwanted heating, especially during consecutive sonications. Automating its delineation with a deep learning algorithm would enhance treatment safety, efficiency and simplify clinical workflow.

The subcutaneous fat layer was manually segmented on MR images from 62 patients treated with MR-HIFU. An attention gated U-Net convolutional neural network (CNN) was trained and validated using the Dice coefficient (DC). Model performance was tested using the DC and 95th percentile Hausdorff distance (HD95). The clinically relevant accuracy was assessed by the average fat layer thickness. The model’s transferability was determined on data prepared by a second reader and compared to the interobserver variability.

The model achieved a DC of 0.972 (IQR: 0.951–0.983), and an HD95 of 1.1 (IQR: 0.8–3.2) mm on the held-out test dataset. The mean absolute thickness error between ground truth and the model’s prediction was 0.8 ± 0.8 mm for the test dataset, and was 0.7–0.8 mm on the two patients prepared by the secondary reader. The automated segmentation algorithm successfully reduced the segmentation time from 3 min to 3 s.

We established an automatic segmentation algorithm based on an attention-gated U-Net architecture to delineate the abdominal fat layer in MR images of uterine fibroids patients. The model achieved high accuracy, robustly handling both thin and thick fat layers, and performed reliable on data prepared by a second reader.