Leveraging Gait Patterns as Biomarkers: An attention-guided Deep Multiple Instance Learning Network for Scoliosis Classification

Haiqing Li, Yuzhi Guo, Feng Jiang, Qifeng Zhou, Hehuan Ma, Junzhou Huang · 2025

Read Leveraging Gait Patterns as Biomarkers: An attention-guided Deep Multiple Instance Learning Network for Scoliosis Classification by Haiqing Li, Yuzhi Guo, Feng Jiang, Qifeng Zhou, Hehuan Ma, Junzhou Huang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-03894.txt.

Introduction

Haiqing Li and co-authors present a 2025 method for spotting scoliosis, a sideways curve of the spine, from how a person walks. Early spotting matters, above all for teens, since a missed curve can squeeze the chest and strain the heart. X-ray checks carry radiation and rely on expert eyes, which limits wide screening. Walking patterns may hold a safer clue.

Their network learns from many short gait clips at once and weighs the most telling ones. It builds on earlier work that first used gait for this task. On the first large gait data set for scoliosis, the method raised accuracy over past tools. It did best on the hard neutral cases, where signs are faint and often missed. It also held up when case types were uneven in count. Walk-based checks could one day screen school halls.

Catalog record

CREATORHaiqing Li, Yuzhi Guo, Feng Jiang, Qifeng Zhou, Hehuan Ma, Junzhou Huang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE19 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-03894

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Leveraging Gait Patterns as Biomarkers: An attention-guided Deep Multiple Instance Learning Network for Scoliosis Classification Haiqing Li, Yuzhi Guo, Feng Jiang, Qifeng Zhou, Hehuan Ma, Junzhou Huang Abstract Scoliosis is a spinal curvature disorder that is difficult to detect early and can compress the chest cavity, impacting respiratory function and cardiac health. Especially for adolescents, delayed detection and treatment result in worsening compression. Traditional scoliosis detection methods heavily rely on clinical expertise, and X-ray imaging poses radiation risks, limiting large-scale early screening. We propose an Attention-Guided Deep Multi-Instance Learning method (Gait-MIL) to effectively capture discriminative features from gait patterns, which is inspired by ScoNet-MT’s pioneering use of gait patterns for scoliosis detection. We evaluate our method on the first large-scale dataset based on gait patterns for scoliosis classification. The results demonstrate that our study improves the performance of using gait as a biomarker for scoliosis detection, significantly enhances detection accuracy for the particularly challenging Neutral cases, where subtle indicators are often overlooked. Our Gait-MIL also performs robustly in imbalanced scenarios, making it a promising tool for large-scale scoliosis screening. † † address: Department of Computer Science and Engineering, The University of Texas at Arlington, Texas, USA 1 Introduction Scoliosis is a spinal curvature disorder that is difficult to detect in its early stages, and for most children and adolescents, the cause is idiopathic [ 1 , 2 ] . Without early detection and timely treatment, scoliosis can progressively worsen, not only affecting the posture but also potentially causing long-term back pain, breathing difficulties, cardiovascular dysfunction, and other issues, severely impacting their health and quality of life. Its diagnostic criteria primarily are assessing the degree of spinal curvature (usually measured by the Cobb angle on an X-ray, as shown in Fig. 1 and the presence of asymmetry in the thorax, shoulders, or waist height [ 3 ] . If the curvature exceeds 10 degrees, scoliosis is diagnosed [ 4 ] . This method not only requires extensive clinical experience but also exposes patients to radiation, which limits the feasibility of large-scale early screening. With the development of deep learning [ 5 ] , researchers have started exploring non-contact and more efficient detection methods, such as back photo analysis [ 6 ] . These methods [ 7 , 8 ] have advanced the application and development of computer vision technology in the medical field, but they still face challenges related to privacy issues and the need for high user cooperation [ 9 ] . \begin{overpic}[cobb.pdf] \put(16.0,-3.0){(a)} \put(75.0,-3.0){(b)} \end{overpic} Figure 1: Illustrations of Cobb angle calculation, used to evaluate scoliosis severity. Figure 2: The overall flowchart of the proposed Gait-MIL. It consists of four parts: (1) K-Means is used to cluster the sampled gait frames into distinct bags, enabling the model to extract specific features from different gait phases, (2) Attention-based pooling is applied to focus on the most informative frames within each bag, and (3) An attention mechanism is used to aggregate bag features into a global representation, (4) The triplet and cross-entropy losses are utilized to drive the training process. To address this challenge, ScoNet-MT [ 10 ] first leverages gait patterns as biomarkers for scoliosis detection and introduce the first large-scale dataset based on gait patterns (Scoliosis1K) for scoliosis classification. ScoNet-MT processes each frame of participant videos by feeding individual silhouette images into an image encoder. The representations of all frames are then pooled and fed into a scoliosis predictor for final classification. However, gait patterns are inherently complex and dynamic, as each sequence consists of multiple frames, and not all frames contribute equally to the final classification. Multiple Instance Learning (MIL) [ 11 , 12 ] has been demonstrated as a favorable option for addressing this issue, especially in situations where only some instances (frames) within a set (sequence) are informative for making a decision. Inspired by prior works [ 10 , 13 ] , we propose an Attention-Guided Deep Multiple Instance Learning Network (Gait-MIL) for scoliosis classification. Specifically, we cluster all frames in a sequence into several categories, which may represent different phases of gait, such as the start, mid-walk, and end, or phases like left foot lift, right foot lift, and both feet on the ground. Each cluster is treated as a "bag of instances", furthermore, we introduce an attention mechanism that assigns greater importance to instances (frames) within each bags that are more likely to exhibit scoliosis-related abnormalities. This enables more effective aggregation of information along the temporal dimension. Subsequently, we employ another attention mechanism to assign greater importance to the bags that are more likely to exhibit scoliosis-related abnormalities, thereby addressing the inherent variability in gait sequences. The combination of MIL and attention allows the model to prioritize the most informative instances (frames) for scoliosis detection, filtering out irrelevant information and reducing the risk of overfitting to unnecessary details. At the same time, this approach makes our method particularly robust to noise and variations in gait sequences, which are common in real-world data collection. In summary, this paper contributes from three key aspects: (1) We introduce Gait-MIL, the first framework that applies Multiple Instance Learning (MIL) to scoliosis detection using gait as a biomarker. This novel approach enables a refined focus on key segments within gait sequences, which are particularly relevant for identifying scoliosis indicators. (2) Our framework employs an attention mechanism to identify and emphasize key frames likely to display scoliosis-related abnormalities. Additionally, by integrating a clustering method, we highlight pivotal moments or clusters within the sequence, thereby enhancing performance by focusing on essential temporal patterns. (3) To evaluate the effectiveness of the proposed Gait-MIL , we conduct extensive experiments on the first large-scale dataset specifically designed for video-based scoliosis classification. The results strongly validate our Gait-MIL’s effectiveness, particularly in enhancing the discrimination of Neutral cases, which are often challenging due to their subtle indicators. 2 Methodology The overview of our proposed Gait-MIL is demonstrated in Fig 2 , and the technical details are introduced as follows. 2.1 Frame-Level Clustering In the task of scoliosis classification, each subject X_{i} contains a sequence of gait frames \{f_{i1},f_{i2},\dots,f_{iT}\} , where i denotes the subject index and T represents the length of the complete gait sequence (number of frames). However, during training, the sampler randomly selects S frames for each subject in a batch to from all available frames of each subject to form the training sequence, i.e., \{f_{ij}\}_{j=1}^{S} , where S\leq T . This sampling introduces complexity and dynamic variability into the sequence, making it impractical to use these sequences as instances directly since the gait frames no longer retain their original temporal order. To address this, K-Means [ 14 ] is introduced to cluster the frames within the sequence into multiple bags (phases) . This clustering method effectively handles the heterogeneity of different phases, allowing the model to extract phase-specific features from the gait data. Therefore, the sampled gait frames f_{ij} are partitioned into multiple bags (phases), i.e., \{b_{i1},b_{i2},\dots,b_{iK}\} , where K represents the number of bags formed after clustering. This clustering approach enables the model to extract representative features from different gait phases. 2.2 Attention-Based Pooling More specifically, even after clustering the frames into different phases, the importance of individual frames varies. To effectively

Read the complete text