Toward automated plantar pressure analysis: machine learning-based segmentation and key point detection across multicenter data

Abstract

Plantar pressure analysis is a pivotal tool for assessing foot function, diagnosing deformities, and characterizing gait patterns. Traditional proportion-based segmentation methods are often limited, particularly for atypical foot structures and low-quality data. Although recent advances in machine learning (ML) offer opportunities for automated and robust segmentation across diverse datasets, existing models primarily rely on data from single laboratories, limiting their applicability to multicenter datasets. Furthermore, the prediction of relevant landmarks on the plantar pressure profile has not been explored. This study addresses these gaps by exploring ML-based approaches for anatomical zone segmentation and landmark detection in plantar pressure analysis, including 758 plantar pressure samples from 460 individuals (197 females, 263 males) collected from multiple centers during static and dynamic conditions using two distinct systems. The datasets were further standardized and augmented. The plantar surface was segmented into four regions (hallux, metatarsal area 1, metatarsal areas 2-5, and the heel) using a U-Net model, and deep learning regression models predicted the key points, such as interdigital space coordinates and the center of metatarsal area 1. The results underscore the U-Net's capacity to attain an accuracy comparable to that of experts (Median Dice Scores ≥ 0.88), particularly in regions with well-defined plantar pressure boundaries. Metatarsal area 1 exhibited unique characteristics because of its ambiguous boundaries, with expert reviews playing a valuable role in enhancing accuracy in critical cases. Using a regression model (Median Euclidean distance = 7.72) or an ensemble model (Median Euclidean distance = 5.26) did not improve calculating the center of metatarsal area 1 directly from the segmentation model (Median Euclidean distance = 4.47). Furthermore, regression-based approaches generated higher errors in key point detection of the interdigital space 2-3 (Median Euclidean distance = 10.06) than in metatarsal area 1 center (Median Euclidean distance = 7.72). These findings emphasize the robustness of the proposed segmentation and key point prediction models across diverse datasets and hardware setups. Overall, the proposed methods facilitate the efficient processing of large, multicenter datasets across diverse hardware setups, significantly reducing the reliance on extensive human labeling, lowering costs, and minimizing subjective bias through ML-driven standardization. Leveraging these strengths, this work introduces a novel framework that integrates multicenter plantar pressure data for both segmentation and landmark detection, offering practical value in clinical and research settings by enabling standardized, automated analyses across varying hardware configurations.

Keywords: U-net; artificial intelligence; biomechanics; deep learning; hallux angle; image segmentation; intelligent systems; zoning.

FIGURE 1

Workflow overview of the study. The segmentation areas and regression key points are listed and described in more detail in Table 1.

FIGURE 2

U-Net model used in the current study. The architecture consists of an encoder and decoder with skip connections. Each convolutional layer uses a 3 × 3 kernel, followed by batch normalization and LeakyReLU activation. The encoder progressively downsamples the input image through Conv2D and MaxPooling2D layers, while the decoder upsamples the feature maps using UpSampling2D and concatenates the corresponding encoder layers. Dropout is applied in the bottleneck layer to prevent overfitting. The final output layer has four channels with a sigmoid activation function.

FIGURE 3

Example of U-Net-based segmentation for a single participant’s foot. Shown are model-generated predictions for each output channel, representing different segmentation areas, are shown before (upper panels) and after (lower panels) applying a threshold of 0.5. The colorbar represents the predicted probability values of the mask, ranging from 0 to 1, with higher values indicating stronger confidence in the predicted segmentation. Note that only model-based annotations are shown in this figure. For a visual comparison of expert annotations and thresholded model predictions overlaid on normalized and rescaled plantar pressure profiles, refer to Figure 4 (upper example).

FIGURE 4

Exemplary normalized and rescaled plantar pressure profile with overlaid segmentation masks: expert annotation (left) and model-predicted, thresholded mask (right; see Figure 3 for thresholding details). True and predicted key points generated using different approaches are also shown (see Section 3.2 for methodological details).

FIGURE 5

Euclidean distance differences between predicted and actual points across all samples are used to estimate the Gaussian kernel density for the center of metatarsal area 1 via segmentation and shifted interdigital space 2–3. The kernel density contours represent the probability density of these errors. An example of plantar pressure distribution is shown to facilitate error comparison relative to foot size. A customized colormap for the plantar pressure distribution was used to improve visibility of the kernel density contours.

FIGURE 6

Exemplary inter-rater reliability of three raters annotating the anatomical landmarks. Semi-transparent filled polygons represent individual annotations, with overlapping regions visualized through increased color density. Solid borders highlight individual outlines. Note: The annotations shown reflect the original shapes as provided by the raters prior to resizing to the model’s input shape, preserving the spatial characteristics of the raw labeling process.