Deep Learning to Detect Triangular Fibrocartilage Complex Injury in Wrist MRI: Retrospective Study with Internal and External Validation

Abstract

Objective: To use deep learning to predict the probability of triangular fibrocartilage complex (TFCC) injury in patients' MRI scans.

Methods: We retrospectively studied medical records over 11 years and 2 months (1 January 2009-29 February 2019), collecting 332 contrast-enhanced hand MRI scans showing TFCC injury (143 scans) or not (189 scans) from a general hospital. We employed two convolutional neural networks with the MRNet (Algorithm 1) and ResNet50 (Algorithm 2) framework for deep learning. Explainable artificial intelligence was used for heatmap analysis. We tested deep learning using an external dataset containing the MRI scans of 12 patients with TFCC injuries and 38 healthy subjects.

Results: In the internal dataset, Algorithm 1 had an AUC of 0.809 (95% confidence interval-CI: 0.670-0.947) for TFCC injury detection as well as an accuracy, sensitivity, and specificity of 75.6% (95% CI: 0.613-0.858), 66.7% (95% CI: 0.438-0.837), and 81.5% (95% CI: 0.633-0.918), respectively, and an F1 score of 0.686. Algorithm 2 had an AUC of 0.871 (95% CI: 0.747-0.995) for TFCC injury detection and an accuracy, sensitivity, and specificity of 90.7% (95% CI: 0.787-0.962), 88.2% (95% CI: 0.664-0.966), and 92.3% (95% CI: 0.763-0.978), respectively, and an F1 score of 0.882. The accuracy, sensitivity, and specificity for radiologist 1 were 88.9, 94.4 and 85.2%, respectively, and for radiologist 2, they were 71.1, 100 and 51.9%, respectively.

Conclusions: A modified MRNet framework enables the detection of TFCC injury and guides accurate diagnosis.

Keywords: data analysis; deep learning; magnetic resonance imaging; retrospective study; triangular fibrocartilage complex.

Figure 1

The local and external datasets for model training, validation, and testing.

Figure 2

(A) An overview of both algorithms used in this study. The topmost data flow shows how Algorithm 1 (ResNet50) processes images along the coronal axis. ResNet50 itself was originally a model for processing 2D images. For ResNet50 to integrate the information from the z-axis, the MR images would first go through the z-direction cropping, which selects the middle eight of a set of MR images. Z-compression then transforms the eight images into three channels. In the ResNet50 architecture, this study uses the model’s channel dimension as the z-dimension. Other data flows show how Algorithm 2 (MRNet) processes images along three axes. Training of MRNet is two-stage. Three AlexNets are optimized respectively, indicated by the three log-likelihood maximization. In the second stage training, their outputs are then passed to a logistic regression classifier for an ensemble result. In the design for the MRNet, where the backbone AlexNet was also designed for 2D images, batch dimension was used for the z-dimension, and information along the z-axis was integrated with z-max pooling. (B)The architecture of ResNet50, the backbone of Algorithm 1 in this study. ResNet50 features residual links, indicated by the jumping arrows to the right of the layer stacks, facilitating the passing of information into a very deep network. Each convolution layer (colored block) is followed by a batch normalization and ReLU activation function, which are not shown in this Figure 2. (C) The architecture of AlexNet, the main backbone of MRNet in this study. The feature extraction of AlexNet consists of five convolution layers and three max-pooling layers, in the depicted order. In the original AlexNet, feature extraction ends with global average pooling to 256 × 6 × 6 and the output tensor were flattened to a 9216-dimensional vector for the downstream classifier. Here, we followed the MRNet’s design and average-pool the output to 256 × 1 × 1 before entering the fully-connected network.

Figure 3

The ROC curves of algorithms (a) 1 and (b) 2 for the internal dataset. The ROC curves of algorithms (c) 1 and (d) 2 for the external dataset. ROC, receiver operating characteristic.

Figure 4

The activation heatmaps overlaid on the original MRI scans. The white arrows indicate the overlapping area of the contrast media and heatmaps.