CurvNet: Latent Contour Representation and Iterative Data Engine for Curvature Angle Estimation

TL;DR

CurvNet presents a novel approach for automatic Cobb angle estimation from scoliosis X-rays by integrating a latent contour representation with an eigen-spine latent space and a regression-plus-classification contour detector. The method applies an iterative data engine that generates realistic, privacy-preserving images and progressively refines annotations through pseudo-labeling, auto-annotation, manual correction, and privacy review, resulting in the large open Spinal-AI2024 dataset. Across public, private, and generated datasets, CurvNet achieves state-of-the-art performance on maximum and regional Cobb angles, while ablation studies demonstrate the critical roles of both the latent contour framework and the data-engineering pipeline. The work advances automatic scoliosis assessment by offering high accuracy, robust generalization across data scales, and a privacy-conscious data-generation paradigm with practical clinical impact.

Abstract

Curvature angle is a quantitative measurement of a curve, in which Cobb angle is customized for spinal curvature. Automatic Cobb angle measurement from X-ray images is crucial for scoliosis screening and diagnosis. However, most existing regression-based and segmentation-based methods struggle with inaccurate spine representations or mask connectivity and fragmentation issues. Besides, landmark-based methods suffer from insufficient training data and annotations. To address these challenges, we propose a novel curvature angle estimation framework named CurvNet including latent contour representation based contour detection and iterative data engine based image self-generation. Specifically, we propose a parameterized spine contour representation in latent space, which enables eigen-spine decomposition and spine contour reconstruction. Latent contour coefficient regression is combined with anchor box classification to solve inaccurate predictions and mask connectivity issues. Moreover, we develop a data engine with image self-generation, automatic annotation, and automatic selection in an iterative manner. By our data engine, we generate a clean dataset named Spinal-AI2024 without privacy leaks, which is the largest released scoliosis X-ray dataset to our knowledge. Extensive experiments on public AASCE2019, our private Spinal2023, and our generated Spinal-AI2024 datasets demonstrate that our method achieves state-of-the-art Cobb angle estimation performance. Our code and Spinal-AI2024 dataset are available at https://github.com/Ernestchenchen/CurvNet and https://github.com/Ernestchenchen/Spinal-AI2024, respectively.

Figures (6)

• Figure 1: Illustration of Cobb angle measurement. There are three categories of scoliosis: proximal thoracic (PT) curve, main thoracic (MT) curve, and thoracolumbar/lumbar (TL/L) curve, as shown in (c). In (a), handwritten 23°, 31°, and 34° are manually measured by experts, angles in green are measured based on rectangular boxes (also in green), and angles in red are measured based on our proposed contour boxes composed by spinal landmarks (drawn in (b)). (b) and (c) show two types of evaluations: maximum angle of three spinal curves, three regional angles, respectively.
• Figure 2: The overall structure of our CurvNet framework, which consists of spinal image generation, spinal contour detection, and data engine. Our private Spinal2023 dataset is used for training spinal image generation model, and also used for training initial model of spinal contour detection network. During each round of data engine, trained spinal contour detection network is adopted for pseudo-labeling, and selected data with annotations are backward used to fine-tune the spinal contour detection network. By employing the data engine to select and annotate generated images, we obtain the Spinal-AI2024 dataset.
• Figure 3: Visualization of different spine representations on example images from AASCE2019 and Spinal2023.
• Figure 4: Performance of our spinal contour detection network under different confidence thresholds on generated dataset obtained at the first data engine iteration. Note that when $\tau_c$ is set optimally at 0.3, ED and SMAPE results on the final Spinal-AI2024 dataset will be increased from 20.73 and 26.13% to 6.23 and 5.34%, respectively.
• Figure 5: Illustration of discarded spinal segment instances as well as unreasonable samples during sample selection. (a) Inaccurate spinal segment instances due to low threshold; (b) Spinal segment instances with area smaller than 200 pixel$^2$; (c) Samples with fewer than 10 instances; (d) Image distortion; (e) Spinal fracture; (f) Samples with excessively high comprehensive similarity (CS).