Source-linked AI summary
Predicting the Progression of Adolescent Idiopathic Scoliosis
Owen Pullen, Amir Jamaludin, Andrew Zisserman
TL;DR
Predicting scoliosis progression has received little research compared with automated severity measurement. This paper predicts spine curves across DXA time points using synthetic pretraining and fine-tuning on real data; the authors report accurate predictions on real sequences and improved performance after fine-tuning.
Problem
Automated scoliosis severity measurement has received attention, but disease-progression prediction has received little research.
Source: §1 · p. 1Method
A transformer predicts DXA spine-curve sequences after synthetic-data pretraining and fine-tuning on real ALSPAC data.
Source: §1 · p. 1, §3.2 · p. 8, §5 · p. 15Results
The model predicts real scoliosis curve sequences, and fine-tuning on real data increases accuracy and consistency over the pretrained model.
Source: §4.3 · p. 14, §5 · p. 15Takeaways & Limitations
The model is intended to predict missing scan time points for research or the next scan for clinical follow-up decisions.
Source: §1 · p. 1Real ALSPAC sequences often have fewer than six scans, and missing time-point locations can make prediction more difficult.
Source: §4 · p. 11
Abstract
from arXiv · showhide
Adolescent Idiopathic Scoliosis is defined as a lateral curvature of the spine that develops during adolescence, without known cause. The condition can result in significant pain and disability, and often progresses rapidly during adolescence. The objective of this paper is to predict the progression of the condition in a temporal sequence from ages 9 to 24, as measured from a sequence of Dual X-ray Absorptiometry (DXA) scans. To this end, we train a transformer model that takes in the curve of the spine to predict curve progression. The model is trained using a large-scale synthetic dataset of spine curves and their time series, covering different curve types and different progression patterns. We show that the model is able to generalise from synthetic to real data by evaluating it on a dataset of real DXA scans covering multiple time points. We find that fine-tuning the model on real data gives a significant boost to performance. The model is able to accurately predict spine curve progression in both scoliosis and normal cases.
1 Introduction
The paper targets a gap in scoliosis research: predicting disease progression has received little attention compared with automated severity measurement. It proposes predicting spine-curve progression from longitudinal DXA scans, using synthetic pretraining followed by fine-tuning on real data.
1 Introduction
- Although automated scoliosis severity measurement has received attention, little research has attempted to predict disease progression.The paper uses ALSPAC as its primary dataset.Source: §1 · p. 1, Figure 1 · p. 2
- The model is intended to predict unobserved DXA scan time points, either to fill gaps in sequences or forecast the next scan.The paper identifies research imputation and clinical follow-up decisions as distinct use cases.Source: §1 · pp. 1–2
- Because real data are scarce, the authors generate synthetic curves and progression sequences to pretrain a transformer, then fine-tune it on the smaller ALSPAC dataset.Pretraining masks time-series samples and predicts them from the observed samples.Source: §1 · pp. 2–3
- The paper’s stated novelty is predicting spinal-shape progression to support estimates of scoliosis severity and clinically relevant phenotypes.The authors also report that their synthetic pretraining dataset boosts performance.Source: §1 · pp. 2–3
1.1 Related Work
- Prior work includes Cobb-angle measurement on X-rays and the DXA-adapted DSM, which uses a straight-spine reference when vertebral endplates are unclear.The DSM measures the angle between lines from the curve’s end vertebrae to its apex.Source: §1.1 · p. 3, Figure 2 · p. 3
- Prior progression-prediction studies considered future Cobb-angle regression or severe disease as a binary outcome, rather than the paper’s DXA curve-sequence task.Related DXA work focused on automated DSM-angle measurement.Source: §1.1 · pp. 3–4
2 Dataset: the ALSPAC study
The ALSPAC dataset provides adolescent DXA scans at six ages, with varied follow-up completeness and progression patterns. The study uses a stratified subset for evaluation and defines a DXA-angle threshold for severe disease.
2 Dataset: the ALSPAC study
- ALSPAC participants were scanned at ages 9, 11, 13, 15, 17, and 24; the dataset table summarizes scan counts by age and participants by number of missing time points.Because scans are supine, DSM angles are lower than typical X-ray Cobb angles; the study uses 16° as its severe-disease threshold.Source: §2 · p. 4, Table 1 · p. 4
- Among individuals with a scan above 16°, most progress gradually or do not reach severe disease rapidly, while a small minority progress rapidly.The sequences are presented as showing each individual’s overall direction of progression.Source: §2 · p. 4, Figure 3 · p. 5
- Evaluation used a randomly sampled, severity-stratified half of the 9,121-person cohort: 4,561 individuals with 16,708 DXA scans.Eight participants with surgical interventions were censored from the time surgery was first recognized; non-surgical bracing information was unavailable.Source: Real evaluation data · pp. 4–5
3 Learning to Predict Progression
The paper uses an encoder-only transformer to predict scoliosis progression from DXA scans, pre-training on synthetic curve sequences and then fine-tuning on real ALSPAC data. The section describes how synthetic curves, model inputs, and angle estimates are constructed for this task.
3.1 Synthetic Data Creation
- Synthetic curves use real-data bounds for peak values and overall change, while normalizing curves to a shared vertical extent.The authors state that synthetic and real distributions need not match, avoiding reproduction of severe-scoliosis sparsity in the synthetic data.Source: §3.1 · pp. 5–6
- The synthetic pipeline creates lumbar, thoracic, thoracolumbar, and normal curves, including left- and right-sided single and compensatory patterns, with four broad progression classes.Progression classes are increasing, plateauing followed by increasing, stable, and stable decrease.Source: §3.1 · pp. 5–6
3.2 Curve Progression Model
- The encoder-only transformer predicts masked time points from six curve tokens, using a 10-coefficient Fourier representation of each normalized spine curve.The Fourier representation is projected to the transformer dimension and projected back to coefficients for curve reconstruction.Source: §3.2 · pp. 7–8
3.3 Training
- Pre-training reconstructs randomly masked synthetic time points with a self-supervised coefficient loss; fine-tuning uses real ALSPAC sequences and combines coefficient and angle losses.The real-data split is by individual, and the same test set evaluates the pretrained and fine-tuned models.Source: §3.3 · pp. 8–9
3.4 DSM Angle Regression Network
- The DSM angle regressor maps the 10-dimensional Fourier representation to an estimated angle and was evaluated on ALSPAC data unseen during its training.It was trained on annotated ALSPAC and UK Biobank data; the reported test metrics are given above.Source: §3.4 · p. 10
- The DSM angle regression network achieved a test MAE of 1.62° and MAPE of 19.28%, with Pearson’s correlation coefficient of 0.80 on ALSPAC data.The Pearson correlation is comparable to [6], which reports 0.82 with an MAE of 1.9°.Source: Figure 7 · p. 10, §3.4 · p. 10
4 Results
This section evaluates interpolation and future-point prediction against curve- and PCA-space baselines. It reports strong model performance, with fine-tuning improving predictions, while noting that real sequences often have missing scans.
4 Results
- Evaluation uses held-out synthetic and ALSPAC data for age-15 interpolation and future-masked extrapolation at ages 15, 17, and 24, scoring the proportion within 2°–5°; real sequences often lack scans.Missing scans at different positions can make prediction more difficult, particularly when a scan just before the target is absent.Source: §4 · p. 11
4.1 Baselines
- Interpolation baselines weight observed curves by inverse age distance; extrapolation baselines fit a direction and per-year rate in curve or PCA space.The PCA representation uses 10 components, retaining 99.5% of variance.Source: §4.1 · pp. 11–12
4.2 Interpolation results
- For scoliosis cases, the model beats the baselines; fine-tuning improves the pre-trained model at 2° and 5° thresholds, while baselines lead for non-cases.The baselines perform well for non-cases because their angles change little across scans.Source: Table 3 · p. 12, §4.2 · p. 13
4.3 Extrapolation results
- The model predicts complex increases robustly against baselines, and fine-tuning improves prediction quality across ages 15, 17, and 24.The extrapolation task masks the target and all later time points to prevent the model from looking ahead.Source: Table 4 · p. 14, §4.3 · pp. 14–15
- ALSPAC-only training performs well for 0–6° angles but less well on moderate and severe cases, where severe scoliosis is rare in the real data.The passage suggests the model may struggle to learn complex increase patterns from these sparse cases.Source: §4.3 · p. 14
5 Conclusion
The paper concludes that a transformer pre-trained on synthetic data and fine-tuned on real data can predict spine curves in real DXA sequences of adolescent idiopathic scoliosis. It identifies potential clinical and research uses, alongside future work on contextual inputs and cross-institute testing.
5 Conclusion
- A transformer pre-trained on synthetic data and fine-tuned on real data can predict spine curves in real DXA sequences of adolescent idiopathic scoliosis.Source: §5 · p. 15
- The authors suggest the model could support follow-up decisions for patients at risk of progression and earlier access to physiotherapy or bracing.Source: §5 · p. 15
- The model could also help researchers impute missing observations; future work includes adding contextual information and testing across institutes.Source: §5 · p. 15