Advances in Ultrasound Imaging: Automated 3D Segmentation of Anatomical Structures in the Forearm

Technical Summary: Automated 3D Segmentation of Anatomical Structures in the Forearm

Problem Statement
Musculoskeletal ultrasound (MSUS) is a cornerstone in rheumatology and clinical medicine, yet achieving proficiency remains a significant barrier due to the substantial training required. While image quality and functional modalities (e.g., power Doppler, elastography) have improved, the interpretation of 2D cross-sectional images relies heavily on the operator's ability to mentally reconstruct 3D anatomy. This cognitive load, combined with the time-consuming and costly nature of traditional training (courses, videos, mentorship), limits the accessibility of high-quality MSUS for less experienced personnel. Furthermore, existing 3D visualization methods for solid organs and specific structures (nerves, muscles) have not been comprehensively applied to the complex anatomy of the lower forearm flexor side in an automated, multi-structure context.

Methodology
The study utilized a dataset comprising ultrasound images from 50 individuals: 25 patients with established carpal tunnel syndrome (CTS) and 25 healthy controls. Scans were performed on the flexor side of the lower forearm using a GE Logiq 10 scanner with a 6–15 MHz probe in MSK mode. The imaging protocol captured video loops from 15 cm proximal to the wrist crease down to the carpal tunnel entrance.

• Data Annotation: Approximately 50 frames per individual were manually segmented by an experienced rheumatologist. Thirteen anatomical structures were targeted, but the study focused on the segmentation of 11 specific structures: the median nerve, radial artery, radial and ulnar bones, scaphoid and pisiform bones, and five muscles (pronator quadratus, flexor carpi radialis, flexor pollicis longus, flexor digitorum profundus, and flexor digitorum superficialis).
• Algorithm Architecture: A fully convolutional encoder–decoder U-net type network was employed. The architecture features six levels with cross-over connections, utilizing 2×2 max pooling in the encoder and 2×2 repeat upsampling in the decoder. Each level includes two 3×3 convolution layers with ReLU activation.
• Training Protocol: Images were exported with minimal filtering, downscaled to 256×256 pixels, and used without contrast augmentation. The model was trained using Keras with a 10-fold cross-validation strategy, an Adam optimizer, 150 epochs, and a Dice loss function. To prevent overfitting, random augmentations were applied, including gamma intensity transformation, random horizontal flipping, and the application of random Gaussian shadows.
• Evaluation Metrics: Performance was assessed using the Dice score for segmentation accuracy and the F1-score for detection accuracy. A detection was considered a true positive if the automated segmentation overlapped by more than 25% with the manual ground truth. Segmentation regions smaller than a 7-pixel radius were ignored.
• Visualization: 3D visualizations were generated by approximating position data based on uniform, linear probe movement. The segmented data and original ultrasound frames were rendered using the CustusX navigation platform, allowing for 2D images to be embedded within 3D anatomical contexts.

Key Results
The algorithm demonstrated varying levels of success across the 11 segmented structures:

• Detection Accuracy: The median nerve and pronator quadratus muscle achieved the highest detection accuracy (0.98). Seven of the eleven structures exceeded a detection accuracy of 0.90. The pisiform bone showed the lowest detection accuracy (0.40).
• Segmentation Accuracy: The highest segmentation accuracy was observed for the flexor digitorum superficialis muscle (0.83), while the pisiform bone had the lowest (0.25).
• Visualization: The study successfully generated 3D visualizations where specific structures (e.g., median nerve in blue, radial artery in red, bones in yellow) could be isolated or viewed in combination with semi-transparent 2D ultrasound overlays.
• Limitations Observed: The authors noted specific failure modes, such as the algorithm misclassifying part of the flexor digitorum superficialis muscle as a second flexor carpi radialis muscle. Additionally, the low Dice scores for the pisiform bone were attributed to the dataset's limited field of view, which often ended at the bone rather than capturing its full extent.

Significance and Claims
The paper claims to be the first work to demonstrate automated segmentation and annotation of the flexor muscles of the lower forearm via ultrasound. The primary significance lies in the potential to transform MSUS from a purely operator-dependent skill into a technology-assisted tool. By shifting the reconstruction of anatomy from the operator's mind to the imaging system, the technology offers:

1. Enhanced Training: A bedside teaching tool that provides anatomically intuitive representations, supporting spatial understanding for less experienced clinicians.
2. Reproducibility: Automated segmentation allows for consistent identification of structures, reducing inter-operator variability.
3. Contextual Visualization: The ability to view 2D cross-sections within a 3D skeletal environment helps bridge the gap between 2D imaging and 3D anatomical reality.

The authors maintain a modest stance, acknowledging that the technology does not yet achieve 100% accuracy and that the current dataset lacks probe tracking, which limits the smoothness of 3D reconstructions. They posit that while fully automated robotic exams may hinder manual skill development, AI-supported real-time correction of probe positioning could improve operator skills and standardize imaging protocols.