[Paper Review] Automated Rib Fracture Detection of Postmortem Computed Tomography Images Using Machine Learning Techniques

Samuel Gunz, Svenja Erne+6 more|arXiv (Cornell University)|Aug 15, 2019

Autopsy Techniques and Outcomes24 references4 citations

TL;DR

This study proposes a novel machine learning approach for automated rib fracture detection in postmortem CT scans by transforming 2D rib unfolding images into 3D coordinate spaces using RGB color mapping and computing HOMFLY polynomial invariants to classify fractures. The method achieves a precision of 0.60, demonstrating feasibility despite challenges in topology consistency and dataset limitations.

ABSTRACT

Imaging techniques is widely used for medical diagnostics. This leads in some cases to a real bottleneck when there is a lack of medical practitioners and the images have to be manually processed. In such a situation there is a need to reduce the amount of manual work by automating part of the analysis. In this article, we investigate the potential of a machine learning algorithm for medical image processing by computing a topological invariant classifier. First, we select retrospectively from our database of postmortem computed tomography images of rib fractures. The images are prepared by applying a rib unfolding tool that flattens the rib cage to form a two-dimensional projection. We compare the results of our analysis with two independent convolutional neural network models. In the case of the neural network model, we obtain an $F_1$ Score of 0.73. To access the performance of our classifier, we compute the relative proportion of images that were not shared between the two classes. We obtain a precision of 0.60 for the images with rib fractures.

Motivation & Objective

• To reduce manual workload in forensic radiology by automating rib fracture detection in postmortem CT scans.
• To explore the application of topological invariants, specifically HOMFLY polynomials, in classifying medical images.
• To evaluate whether polynomial knot invariants derived from image convolutions can distinguish rib fractures from non-fractures.
• To compare the performance of the topological method against deep learning models (MobileNet and simplified CNN) on the same dataset.
• To identify limitations in current image preprocessing and topology generation for medical imaging applications.

Proposed method

• Rib cages from postmortem CT scans were unfolded into 2D projections using a rib unfolding tool.
• Grayscale CT images were converted to RGB using the nipy_spectral colormap to enable multi-channel processing.
• 13 convolutional kernels were applied separately to each RGB channel, followed by pooling to generate 2×2 feature maps.
• Feature map coordinates (xR, xG, xB) were extracted and mapped into a 3D Euclidean space representing the RGB color space.
• HOMFLY polynomials were computed for each image’s coordinate set to serve as topological invariants for classification.
• Polynomials unique to each class (fracture vs. no fracture) were used to compute precision by measuring exclusive presence in one list.

Experimental results

Research questions

• RQ1 Can HOMFLY polynomial invariants derived from image convolutions reliably classify rib fractures in postmortem CT scans?
• RQ2 How does the performance of the topological classifier compare to established deep learning models like MobileNet and a simplified CNN?
• RQ3 What are the limitations of using kernel-based convolution and topological invariants for medical image classification, particularly in low-contrast, complex anatomical structures?
• RQ4 To what extent does dataset size affect the consistency and accuracy of topological feature generation in this context?
• RQ5 Can a topological approach based on coordinate clustering and polynomial invariants achieve reliable classification despite multiple knot types per object?

Key findings

• The topological classifier achieved a precision of 0.60 in distinguishing rib fractures from non-fractures based on exclusive polynomial presence.
• The F1 score for the two deep learning models (MobileNet and simplified CNN) was 0.73, outperforming the topological method.
• Despite the method's conceptual novelty, it produced multiple distinct topologies for the same object, indicating inconsistency in feature extraction.
• The topological approach struggled with overlapping topologies across classes, limiting its predictive power on unlabelled data.
• The performance was not significantly affected by model architecture, as both MobileNet and the simplified CNN yielded comparable results.
• The authors suggest that medical images with blurred or indistinct contours may not be well-suited for standard deep learning kernels, warranting further investigation into specialized kernel design.