Explainable vertebral fracture analysis with uncertainty estimation using differentiable rule-based classification

TL;DR

A novel method for explainable vertebral fracture assessment (XVFA) in low-dose radiographs using deep neural networks, incorporating Genant's semi-quantitative criteria as a differentiable rule-based means of classifying both vertebra fracture grade and morphology.

Abstract

We present a novel method for explainable vertebral fracture assessment (XVFA) in low-dose radiographs using deep neural networks, incorporating vertebra detection and keypoint localization with uncertainty estimates. We incorporate Genant's semi-quantitative criteria as a differentiable rule-based means of classifying both vertebra fracture grade and morphology. Unlike previous work, XVFA provides explainable classifications relatable to current clinical methodology, as well as uncertainty estimations, while at the same time surpassing state-of-the art methods with a vertebra-level sensitivity of 93% and end-to-end AUC of 97% in a challenging setting. Moreover, we compare intra-reader agreement with model uncertainty estimates, with model reliability on par with human annotators.

Figures (4)

• Figure 1: Visualization of GSQ. Vertebral fracture morphology in Genant's system, with landmarks $\boldsymbol{x}$ (dots). The posterior, middle, and anterior heights (lines) are used to determine deformation morphology and severity. Fig. adapted from Genant1993.
• Figure 2: Model overview. A detector $d_{\theta'}$ is trained to estimate bounding boxes (red) from vertebra in a spinal image (green). Ground truth crops with keypoints (green) are separately used to train a keypoint regressor $f_{x,\theta}$, subsequent vertebra classifiers $f_{img, \theta}, f_{kps}$, as well as estimate the keypoint distribution (red points).
• Figure 3: Visualizations. (a) Test sample model prediction likelihoods (red) and ground truth keypoints (green). GSQ decision borders for (b) morphology and (c) grade, for visual explanation of model results, each point corresponding to a predicted fractured vertebra in the test set.
• Figure 4: Confusion matrices. Confusion matrices for the classification of each severity grade and morphology class, normalized by true classes.