Deep Sequential Learning for Cervical Spine Fracture Detection on Computed Tomography Imaging

Each pixel $x_{i}\in\mathbf{x}_{n}$ is a quantitative representation of the radiodensity of different substances in the scanning area where $-1,000\lesssim x_{i}\lesssim 3,000$, [7]. Fractures of the cervical spine are heterogeneous in their location, type, and composition. They are defined as a break in at least one bone in the cervical spine. The cervical spine can be malaligned as a result of a fracture and/or a ligamentous injury. Both fractures and ligaments injuries can be associated with localized bleeding and soft tissue swelling. In consultation with a panel of radiologists, three window width and center schemes are chosen to enhance cervical spine bone and surrounding tissues, which are soft tissue window $(w_{1}=300;c_{1}=80)$, standard bone window $(w_{2}=1,800;c_{2}=500)$, and gross bone window $(w_{3}=650;c_{3}=400)$. Figure 2 shows the corresponding functions to the proposed windowing schemes to map $x_{i}$ to three different values $\tilde{x}_{i}^{(1)}$, $\tilde{x}_{i}^{(2)}$, $\tilde{x}_{i}^{(3)}$. The represented image with the gross bone window $\tilde{\mathbf{x}}_{n}^{(3)}$ is then used to crop the images with $5\%$ margin from each side of detected cervical spine using Otsu’s method [8]. The cropped images are normalized and resized (padded with zero if not square) to $384\times 384$.

For the image $\mathbf{x}_{n}\in\mathbf{X}$, the preprocessed images are concatenated as ${\tilde{\mathbf{x}}_{n}=(\tilde{\mathbf{x}}_{n}^{(1)}\oplus\tilde{\mathbf{x}}_{n}^{(2)}\oplus\tilde{\mathbf{x}}_{n}^{(3)})}$ where $\oplus$ is the concatenation operator. The cervical spine images have spatiotemporal dependency. First, ResNet-50 [9] is trained using randomly selected batches of images from all training cases. The objective is to train ResNet-50 as the function $\mathbf{F}=\phi(\tilde{\mathbf{X}})$ where $\mathbf{f}_{n}\in\mathbf{F}$ is the extracted feature map from $\tilde{\mathbf{X}}_{n}\in\mathbf{X}$. As Figure 1 shows, the feature maps are vectorized and concatenated as $\tilde{\mathbf{f}}=(\mathbf{f}_{1}\oplus\mathbf{f}_{2}\oplus,...,\mathbf{f}_{N})$. The second phase is learning the temporal dependency among axial images using a bidirectional network of LSTM units to map $\tilde{\mathbf{f}}$ to the case label $y$. The loss is calculated using binary cross-entropy function with respect to the target label $y$. In the inference mode, the feature extractor $\phi(\cdot)$ and BLSTM layer work together for a given case $\mathbf{X}$ to generate the inference $\tilde{y}$.

The dataset contains cervical spine CT scans of 3,666 cases (729 positive and 2,937 negative cases) which corresponds to 1,174,335 images (20,392 fracture positive and 1,153,943 fracture negative). This dataset is pixel-level annotated by three expert radiologists and the intersection of annotated regions was reviewed by a neuroradiologist. Table 1 shows the distribution of gender and age of cases over the fracture positive and negative classes. This tables shows that from a gender perspective, men are more prone to cervical spine fractures. From an age perspective, the average of positive cases is 9 years greater than negative cases which suggests this injury is more common among the elderly [11].

Training setup of the BLSTM [12] model was as follows: Adam optimizer [13] with a learning rate of $10^{-6}$; 100 training epochs; Batch size of 4; Input size of 2,048. Setup of the ResNet-50 [3] model was as follows: Adam optimizer [13] with a learning rate of $10^{-3}$ and step decay of 5 with gamma 0.2; Number of input channels was 3; Random horizontal flip and rotation (10 degrees) augmentation method; Batch size of 16 and 50 training epochs. Early stopping and dropout regularization methods were applied. Given above setup, the proposed model was trained from scratch with randomly initialized weights and tested using a 7-fold cross-validation scheme.

Figure 3 shows two samples of axial cervical spine images with corresponding ground-truth masks and heatmaps generated using Grad-CAM [10] from the last layers of ResNet-50. The heatmaps show ResNet-50 can capture most fracture areas with a relatively high false positive rate. Classification performance of the ResNet-50 at image level fracture detection in cervical spine axial images is presented in Table 2. In this experiment, cervical spine fracture is predicted just based on the spatial features in each axial image without counting temporal information. The results show $80.01\%$ and $77.61\%$ classification accuracy for the imbalanced and balanced datasets, respectively. However, the main drawback of this approach is the high level of false positives which leads to an inaccurate prediction for the entire case. Such that even if a single image in a case is false positive, the entire case will be false positive and vice-versa. These results show the importance of incorporating temporal features in training and inference. Table 3 shows performance results of the combined BLSTM and ResNet-50 model at the case level for different number of LSTM units. The results regarding the imbalanced dataset show as the number of units increases from $96$ to $256$, the classification accuracy also increases from $75.79\%$ to $79.18\%$. However, the accuracy is approximately $71\%$ for the balanced dataset and it is less dependent on the number of LSTM units.

Cervical spine imaging cases from a fracture detection perspective are highly biased from two perspectives. First is the larger number of negative cases when compared to positive cases. The second is generally a greater number of negative images than positive images within a positive case. These results clearly show the importance of considering these biases. For the imbalanced dataset, which reflects the natural distribution of negative and positive cases, as more LSTM units are utilized the imbalanced accuracy also increases. However, we observed no significant change in balanced accuracy. The higher performance on the imbalanced dataset is mainly due to the bias of the dataset toward negative cases and images and bias of the BLSTM layer in capturing the dependency between negative cases.