Automated Segmentation of Intrinsic Foot Muscles in Ultrasound Using Deep Neural Networks: Development and Validation...

To develop and validate the first automated deep-learning image-processing framework for pixel-level segmentation of the plantar intrinsic foot muscles (PIMs) in ultrasound (US) images. This approach supports objective and reproducible US-based muscle segmentation from conventional B-mode images.

US images of four PIMs (abductor hallucis, flexor digitorum brevis (FDB), quadratus plantae [QP] and abductor digiti minimi) were collected from 200 healthy participants (n = 1099 images). Manual segmentations were used as ground truth. Three deep neural network architectures (U-Net, Feature Pyramid Network [FPN] and DeepLabV3+) with a Mix Transformer backbone were trained and evaluated. Model performance was assessed using the Intersection over Union (IoU) and Dice Similarity Coefficient (DSC), two complementary pixel-level segmentation metrics. Agreement with manual measurements was evaluated using intraclass correlation coefficients (ICC), mean difference and standard error of measurement.

All models demonstrated excellent segmentation performance, with mean IoU values ≥0.90 and mean DSC values ≥0.95 across all muscles. The highest accuracy was observed for the FDB (IoU = 0.95; DSC = 0.97), while QP showed slightly lower performance (IoU = 0.90; DSC = 0.95). FPN and DeepLabV3+ exhibited greater stability (SD: ±0.02-0.03) compared to U-Net (up to ±0.08). Agreement with manual segmentation was high for cross-sectional area (ICC: 0.836-0.960) and moderate-to-high for border length (ICC: 0.705-0.944), with minimal systematic bias. Mean model inference time was under 0.01 s per image, reflecting forward-pass model inference only and excluding image loading, resizing and post-processing.

This study presents the first automated segmentation framework for the PIMs in US. Built on established segmentation architectures with a Mix Transformer backbone, it enables accurate and reliable pixel-level segmentation in under 0.01 s per image; its novelty lies in extending automated musculoskeletal US analysis to this previously unaddressed muscle group. By automating tissue boundary detection at the pixel level, this framework advances US imaging beyond conventional manual measurement, enabling objective and scalable musculoskeletal assessment with reduced operator dependence at the segmentation stage. These findings establish the feasibility of accurate and reproducible PIM segmentation in healthy adults and provide the necessary foundation for subsequent validation in clinical populations.