Detecting pelvic fracture on 3D-CT using deep convolutional neural networks with multi-orientated slab images

Abstract

Pelvic fracture is one of the leading causes of death in the elderly, carrying a high risk of death within 1 year of fracture. This study proposes an automated method to detect pelvic fractures on 3-dimensional computed tomography (3D-CT). Deep convolutional neural networks (DCNNs) have been used for lesion detection on 2D and 3D medical images. However, training a DCNN directly using 3D images is complicated, computationally costly, and requires large amounts of training data. We propose a method that evaluates multiple, 2D, real-time object detection systems (YOLOv3 models) in parallel, in which each YOLOv3 model is trained using differently orientated 2D slab images reconstructed from 3D-CT. We assume that an appropriate reconstruction orientation would exist to optimally characterize image features of bone fractures on 3D-CT. Multiple YOLOv3 models in parallel detect 2D fracture candidates in different orientations simultaneously. The 3D fracture region is then obtained by integrating the 2D fracture candidates. The proposed method was validated in 93 subjects with bone fractures. Area under the curve (AUC) was 0.824, with 0.805 recall and 0.907 precision. The AUC with a single orientation was 0.652. This method was then applied to 112 subjects without bone fractures to evaluate over-detection. The proposed method successfully detected no bone fractures in all except 4 non-fracture subjects (96.4%).

Figure 1

Conceptual diagram of the proposed method. (a) A series of axial CT images obtained from a subject. Each image represents 50 × 50-mm area for easy understanding. (b) Nine synthesized, orientated 2.5D images. Three slab images with thicknesses of 18.6 mm, 9.0 mm, and 0.6 mm are visualized by R-G-B colors, respectively. (c) Detection of 2D fracture candidates. (d) Thickening of 2D fracture candidates. (e) Fracture region detection.

Figure 2

Estimated degree of fracture on multiplanar reconstruction images. Top: axial image; bottom-left: coronal image; bottom-right: sagittal image. L: left; R: right; A: anterior; P: posterior; S: superior; I: inferior. (a) Integrated 3D fracture candidate region overlapping on CT images ($C_{\mathit{th}}$: 0.2). Yellow represents the degree of fracture. (b) Resultant 3D fracture region ($I_{\mathit{th}}$: 6). Yellow represents the detected region. The enlarged image shows raw CT images for the detected fracture region.

Figure 3

Precision-recall curve.

Figure 4

The 3D visualization of fractures. (a) Ground truth fractures. (b) Automatically detected fractures.

Figure 5

Performance dependency on analysis parameters. (a) Relationship between $I_{\mathit{th}}$ and $C_{\mathit{th}}$. (b) Cumulative histogram of IoU of the detected fractures.

Figure 6

Types of fractures. (a) Completely displaced fracture (F1). (b) Incompletely displaced fracture (F2). (c) Compression fracture (F3). Fractures are indicated by triangles.

Figure 7

A 2.5D representation. (a) Image with 31-image thickness. (b) Image with 15-image thickness. (c) Image with 1-image thickness.

Figure 8

The 3D annotation method. (a) Annotated 3D bone surface data. (b) Annotation of completely displaced fracture (F1). (c) Annotation of incompletely displaced fracture (F2) or compression fracture (F3).