Semi-Supervised Learning for Bone Mineral Density Estimation in Hip X-ray Images

Abstract

Bone mineral density (BMD) is a clinically critical indicator of osteoporosis, usually measured by dual-energy X-ray absorptiometry (DEXA). Due to the limited accessibility of DEXA machines and examinations, osteoporosis is often under-diagnosed and under-treated, leading to increased fragility fracture risks. Thus it is highly desirable to obtain BMDs with alternative cost-effective and more accessible medical imaging examinations such as X-ray plain films. In this work, we formulate the BMD estimation from plain hip X-ray images as a regression problem. Specifically, we propose a new semi-supervised self-training algorithm to train the BMD regression model using images coupled with DEXA measured BMDs and unlabeled images with pseudo BMDs. Pseudo BMDs are generated and refined iteratively for unlabeled images during selftraining. We also present a novel adaptive triplet loss to improve the model's regression accuracy. On an in-house dataset of 1,090 images (819 unique patients), our BMD estimation method achieves a high Pearson correlation coefficient of 0.8805 to ground-truth BMDs. It offers good feasibility to use the more accessible and cheaper X-ray imaging for opportunistic osteoporosis screening.

Adaptive Contrast for Image Regression in Computer-Aided Disease Assessment

“…Distance metric learning has been successfully used in regression problems by making the separation margins adaptive. The adaptive triplet loss, for example, imposes larger margins on samples pairs with larger differences in values [17,52]. However, these functions still rely on triplet sampling and only applies constraints within each sampled triplet group.…”

“…By performing BMD regression and feature representation learning with AdaCon in a multi-task approach, we can outperform alternative methods by ensuring better features are learnt, as seen in Table II. Notably, Zheng et al [17] use the adaptive triplet loss function in addition to regression loss for BMD estimation. We show analytically and empirically that contrastive learning with AdaCon is a better alternative to their methodology.…”

“…Adaptive methods for distance metric learning have been used in [17,52] for regression tasks by introducing variable separation margins based on label differences into the standard triplet loss function. The major shortfall of triplet loss however is that it is highly dependant on the triplets sampled, but the number of possible triplets grows at a cubic rate relative to batch size, making it difficult to do so effectively [49][50][51].…”

“…Dual-energy X-ray absorptiometry (DXA) is the gold standard for diagnosing osteoporosis and monitoring changes in BMD over time. We use the BMD readings from DXA scans as ground truth labels and collect corresponding X-ray images taken within 6 months of the DXA scans as our model input, following similar settings in [17,20]. In total, we collect 475 valid hip X-ray images from 317 patients who have had a DXA scan performed between 2016-2020.…”

“…Similar techniques have also been applied for computer vision tasks using natural images, such as facial age estimation [12,13] and single-image depth estimation [14]. Recent progress in image regression has increasingly featured deep learning techniques [15][16][17]. For instance, Luo et al [15] used a VGG backbone [18] on cardiac magnetic resonance volumes to directly regress ventricular volumes.…”