Automated segmentation and classification of lumbar transverse ultrasound views using a two-stage deep learning...

Ultrasound guidance is widely used in lumbar regional anesthesia and chronic pain management because it provides radiation-free, portable, and real-time visualization. Among lumbar ultrasound views, the paramedian transverse view at the transverse process (PTV-TP) and the paramedian transverse foraminal view (PTFV) are essential for procedures such as lumbar plexus blocks and transforaminal injections. However, reliable recognition of these views remains technically challenging, particularly for novice clinicians.

This study aimed to develop and evaluate a deep learning-based two-stage framework for automated detection and classification of PTV-TP and PTFV in lumbar ultrasound images.

Ultrasound images were collected from 425 patients undergoing pain treatment at Peking Union Medical College Hospital. The proposed framework consisted of two stages. First, a U-Net-based semantic segmentation model was used to identify key lumbar anatomical landmarks, namely the transverse process and intervertebral foramen, which were grouped into a single class to reduce annotation complexity and improve robustness. Second, coordinate-based spatial features were extracted from the segmentation masks and classified using a support vector machine (SVM) to distinguish between PTV-TP and PTFV views. An end-to-end CNN-based classifier trained under the same experimental protocol served as a comparison. Segmentation performance was evaluated using the intersection-over-union (IoU) metric, and classification performance was assessed using accuracy.

The proposed U-Net-based segmentation achieved an IoU of 83.21%, outperforming conventional multi-class U-Net segmentation (77.33%). Based on the segmentation results, the coordinate-based SVM classifier achieved an accuracy of 97.66% in differentiating PTV-TP and PTFV views, which was substantially higher than that of the end-to-end CNN classifier (85.00%). Visualization results demonstrated stable performance across representative ultrasound images, including challenging imaging conditions.

The proposed two-stage framework, which integrates semantic segmentation with coordinate-based classification, enables accurate automated recognition of lumbar ultrasound views. By explicitly modeling key anatomical landmarks and their spatial relationships, this approach reduces reliance on operator experience and may support safer and more efficient ultrasound-guided lumbar procedures. Future work will focus on expanding the dataset and validating the generalizability of the method across diverse clinical settings.