Evaluation of AI-based spine segmentation (SpS) for the analysis of lumbar spine dual energy X-ray absorptiometry scans

Abstract

Purpose/Aims

The aim of this study was to evaluate a new deep-learning artificial intelligence (AI) -based model for automated SpS. First, we compared bone mineral density (BMD), trabecular bone score (TBS) and bone surface area outcomes across three methods for SpS: 1) the manufacturer default, 2) the clinical DXA expert (criterion) and 3) the new AI-application. Second, we examined longitudinal reproducibility for the measurement of spine surface area.

Rationale/Background

The antero-posterior (AP) lumbar spine dual energy X-ray absorptiometry (DXA) scan is an important diagnostic measure, used for the assessment of osteoporosis. The quality of the scan is dependent on the accuracy of the vertebral bone mask, derived from bone edge detection and spine segmentation (SpS).

Reducing technical error requires manual validation of the default bone mask for each scan. However, this can be time-consuming in practice.

Methods

A sub-sample of 130 women (mean age: 67.1; BMI: 25.2; with no vertebral anomalies) were selected from the OsteoLaus population cohort, having previously received two LS DXA scans (GE Lunar iDXA, encore v 18), 2.5 years apart. Scans were analyzed according to each of the three methods (default, clinical expert and AI), and the primary outcomes (BMD, TBS and surface area) were compared using Student's t-tests and one-way repeated measures-ANOVA. The coefficient of variation (CV%) for bone surface area was also computed.

Results

There were significant differences in mean BMD and TBS outcomes derived from the default bone mask method compared to the DXA clinical expert (p=0.01, Table 1). There were no differences in BMD and TBS derived using the AI SpS bone mask method compared to the DXA clinical expert (p=0.67, Table 1).

Reproducibility for bone surface area was superior for the clinical expert and the AI model compared to the default method (Table 2).

Implications

The AI based model demonstrated improved accuracy and reproducibility for lumbar spine bone segmentation compared to the default analysis method, and in close agreement with the clinical criterion. Overall, these results suggest that the new AI-based model for automated SpS may be a valuable tool for reducing time and improving accuracy for the analysis of lumbar spine DXA scans.