Heat residual U-Net: segmenting anatomical tissue layers depicted on 3D ultrasound images for cLBP

Chronic low back pain (cLBP) is a prevalent public health issue, affecting about 60% of adults in the United States. Among the imaging choices (e.g., CT, MRI, and ultrasound) for cLBP, three-dimensional (3-D) ultrasound (US) is emerging as a novel tool for assessing cLBP due to its safety, portability and cost-effectiveness. However, the image quality of US is low and cannot provide the same level of detail as CT or MRI for characterizing anatomical structures. Additionally, a single 3-D US scan contains more than one hundred image slices, making it time-consuming to manually identify the anatomical structures depicted on US images. We developed and validated a deep learning segmentation solution, Heat Residual U-Net (HR-UNet), to automatically identify six anatomical layers depicted on the 2-D slices of the 3-D ultrasound images acquired from participants in an ongoing cLBP study. These six layers include dermis, superficial fat, superficial fascial membrane, deep fat, deep fascial membrane, and muscle. In architecture, HR-UNet integrates the preprocessing pipeline of nnU-Net, a heat conduction module, and skip residual connections to improve segmentation accuracy. Our experiments on a B-mode 3-D ultrasound dataset showed that the HR-UNet model achieved a mean Dice score of 82.9% (+1.4%, p-value < 0.05 compared to other methods) across the six anatomical layers, outperforming several state-of-the-art methods.