Deep learning system for automated detection of posterior ligamentous complex injury in patients with thoracolumbar fracture on MRI

Abstract

This study aimed to develop a deep learning (DL) algorithm for automated detection and localization of posterior ligamentous complex (PLC) injury in patients with acute thoracolumbar (TL) fracture on magnetic resonance imaging (MRI) and evaluate its diagnostic performance. In this retrospective multicenter study, using midline sagittal T2-weighted image with fracture (± PLC injury), a training dataset and internal and external validation sets of 300, 100, and 100 patients, were constructed with equal numbers of injured and normal PLCs. The DL algorithm was developed through two steps (Attention U-net and Inception-ResNet-V2). We evaluate the diagnostic performance for PLC injury between the DL algorithm and radiologists with different levels of experience. The area under the curves (AUCs) generated by the DL algorithm were 0.928, 0.916 for internal and external validations, and by two radiologists for observer performance test were 0.930, 0.830, respectively. Although no significant difference was found in diagnosing PLC injury between the DL algorithm and radiologists, the DL algorithm exhibited a trend of higher AUC than the radiology trainee. Notably, the radiology trainee's diagnostic performance significantly improved with DL algorithm assistance. Therefore, the DL algorithm exhibited high diagnostic performance in detecting PLC injuries in acute TL fractures.

Figure 1

Flow chart for selection of the study population (PLC, posterior ligamentous complex; TL, thoracolumbar).

Figure 2

The process of deep learning algorithm system, which consists of the two algorithms Attention U-net and Inception-ResNet-V2. Imaging preprocessing was performed by zero padding, resize, histogram equalization, Min–Max normalization, and gamma correction. After segmentation of fractured vertebral body and background soft tissue anatomy by trained Attention U-net, post-processing of predicted segmentation (label) and patch extraction were performed. Final classification task was performed by Inception-ResNet-V2.

Figure 3

Schematic architectures of the first step deep learning algorithm (Attention U-net) and second step deep learning algorithm (Inception-ResNet-V2). (a) A block diagram of the Attention U-Net segmentation model in this study. The input image is gradually filtered and down-sampled at each step in the encoding portion of the network. In addition to the basic structure of U-Net, where Encoder for obtaining overall context information of the image and Decoder for accurate localization are symmetrically configured, the accuracy is improved by emphasizing only the necessary features using Attention Gates (AGs) for each skip connection. (b) The overall scheme of the Inception-ResNet-v2 networks in this study. As a first step, it goes through a Stem block with a general Convolution and Pooling structure. In the second step, it goes through a combination of Inception-ResNet Block, which combines Inception's features and ResNet's strengths, and Reduction Block, which generates size changes in features. Finally, the probability value of the class is extracted by making the feature into a one-dimensional vector through Global Average Pooling.

Figure 4

ROC plots for deep learning algorithm and radiologists in the external validation dataset. AUROC of the DL-algorithm was not significantly different from that of experienced musculoskeletal radiologist (R1, p = 0.722), but tended to be higher than that of radiology trainee (R2, p = 0.051) on Delong's test, close to statistical significance. There was significantly different in the AUROC between R1 and R2 (p = 0.011). In the second session with deep learning algorithm assistance, significant improvement in diagnostic performance was observed in R2 (increment of AUROC was 0.090, p = 0.007). (ROC, the receiver operating characteristics; AUROC, the area under the curve of the receiver operating characteristic; DL, deep learning).

Figure 5

Representative four cases of FS sagittal T2 weighted image via gradient-weighted class activation mapping. (a,c) TL fracture cases with normal PLC. (b,d) TL fracture cased with injured PLC. Compared to injured PLC cases (b,d) and normal PLC cases (a,c), it can be seen that Grad-CAM is more limited to the background soft tissue anatomy area including injured PLC. This can be interpreted as the DL algorithm making a judgment based on the area where the PLC is injured. (FS, fat suppression; TL, thoracolumbar; PLC, posterior ligamentous complex; DL, deep learning).

Figure 6

Comparison of Grad-CAM results for a single classification algorithm (left column) and the proposed DL system (right column). (a) and (c) are FS T2 sagittal images of two different TL fracture patients with PLC injury. The GRAD-CAM results for Single classification algorithm are (a) and (c), and the GRAD-CAM results for the proposed DL system of (a) and (c) images are (b) and (d), respectively. The yellow dashed boxes indicates the final image patches resulting from the segmentation task and patch extraction. As shown in the figure, the single classification algorithm (a,c) makes predictions based on different locations than the final image patches extracted by patch extraction, such as the yellow boxes, unlike the proposed DL system (b,d). This disparity explains why the AUROC of the single classification algorithm is much lower than that of the proposed DL system (FS, fat suppression; TL, thoracolumbar; DL, deep learning).