Automated Thoracolumbar Stump Rib Detection and Analysis in a Large CT Cohort

TL;DR

This work tackles automated detection and analysis of thoracolumbar stump ribs to support thoracolumbar transitional vertebrae assessment. It presents an end-to-end pipeline combining high-resolution rib segmentation (nnUNet), a novel Rib Length Measurement Algorithm, and morphology-based features to differentiate stump ribs from normal ribs, even with partial field-of-view. The approach achieves near-perfect segmentation accuracy (Dice ≈ $0.997$) and a $98.2$ success rate for rib length measurement, identifying $159$ stump ribs across $648$ subjects and enabling $F1$-scores up to $0.84$ in classification using early rib segments. The method outperforms existing baselines, and the authors publicly release code, weights, and segmentation masks to advance reproducibility and clinical adoption, potentially improving surgical planning and vertebral labeling.

Abstract

Thoracolumbar stump ribs are one of the essential indicators of thoracolumbar transitional vertebrae or enumeration anomalies. While some studies manually assess these anomalies and describe the ribs qualitatively, this study aims to automate thoracolumbar stump rib detection and analyze their morphology quantitatively. To this end, we train a high-resolution deep-learning model for rib segmentation and show significant improvements compared to existing models (Dice score 0.997 vs. 0.779, p-value < 0.01). In addition, we use an iterative algorithm and piece-wise linear interpolation to assess the length of the ribs, showing a success rate of 98.2%. When analyzing morphological features, we show that stump ribs articulate more posteriorly at the vertebrae (-19.2 +- 3.8 vs -13.8 +- 2.5, p-value < 0.01), are thinner (260.6 +- 103.4 vs. 563.6 +- 127.1, p-value < 0.01), and are oriented more downwards and sideways within the first centimeters in contrast to full-length ribs. We show that with partially visible ribs, these features can achieve an F1-score of 0.84 in differentiating stump ribs from regular ones. We publish the model weights and masks for public use.

Figures (9)

• Figure 1: The flow of our three utilized datasets. We used the private in-house dataset for training the rib segmentation only and performed all downstream experiments on public data only.
• Figure 2: Two example subjects (rows) of the RibFrac dataset. We compare the original RibSeg annotation with the prediction of the TotalSegmentator and our rib segmentation model. Both the RibSeg annotation and the one from TotalSegmentator fail to annotate the rib part close to the vertebra. Furthermore, the RibSeg annotation is not dense but contains only the outline, and the one from TotalSegmentator lacks detail as it segments with an isotropic resolution of 1.5 mm. Our model is trained for 0.8 mm isotropic space; thus, it produces smooth and high-resolution annotations. This behavior is true for all samples in our datasets.
• Figure 3: An example showing our transition from semantic rib segmentation to instance rib segmentation by using the vertebra instance segmentation provided by SpineR. From left to right: Binary rib segmentation; vertebra instance segmentation; resulting combined instance annotation.
• Figure 4: 2D Schematic illustration of our 3D rib length measurement algorithm (RLMA). The blue contour represents the rib, while the corresponding vertebra is orange (the vertebral corpus is darker). The first row shows the initialization of the starting point. First, we find the closest point of the rib surface to the center of the corresponding vertebra corpus (a); second, we use the surrounding points to find an average starting position (b), and third, we finalize the starting point by projecting back to the rib surface (c). The second row shows an iteration: Finding candidate points by looking at a fixed circular distance (d), then taking the average of those candidate points and finding the next point on the path by using the direction vector of the previous point and the average of candidate points (e). The last panel (f) shows the example after the RLMA has calculated all path points. The figure demonstrates the algorithm in two dimensions for clarification purposes, while the RLMA works in three-dimensional data.
• Figure 5: Example 2D image to showcase some of our calculated 3D features. Blue is the rib, and orange is its corresponding vertebra. DRC is the spatial distance and direction between the rib start point and the center of the vertebra corpus region (dark orange). PDRC is the posterior component of the DRC. $2$-PPR: The spatial relation for the first two path points (i.e., the direction vector between the start point and the next path point).