A fast and fully-automated deep-learning approach for accurate hemorrhage segmentation and volume quantification in non-contrast whole-head CT

Abstract

This project aimed to develop and evaluate a fast and fully-automated deep-learning method applying convolutional neural networks with deep supervision (CNN-DS) for accurate hematoma segmentation and volume quantification in computed tomography (CT) scans. Non-contrast whole-head CT scans of 55 patients with hemorrhagic stroke were used. Individual scans were standardized to 64 axial slices of 128 × 128 voxels. Each voxel was annotated independently by experienced raters, generating a binary label of hematoma versus normal brain tissue based on majority voting. The dataset was split randomly into training (n = 45) and testing (n = 10) subsets. A CNN-DS model was built applying the training data and examined using the testing data. Performance of the CNN-DS solution was compared with three previously established methods. The CNN-DS achieved a Dice coefficient score of 0.84 ± 0.06 and recall of 0.83 ± 0.07, higher than patch-wise U-Net (< 0.76). CNN-DS average running time of 0.74 ± 0.07 s was faster than PItcHPERFeCT (> 1412 s) and slice-based U-Net (> 12 s). Comparable interrater agreement rates were observed between "method-human" vs. "human-human" (Cohen's kappa coefficients > 0.82). The fully automated CNN-DS approach demonstrated expert-level accuracy in fast segmentation and quantification of hematoma, substantially improving over previous methods. Further research is warranted to test the CNN-DS solution as a software tool in clinical settings for effective stroke management.

Figure 1

Disagreement percentages between each pair of raters. E1, E2, E3 represents expert 1, 2, and 3, respectively, while M indicates the CNN-DS method. Disagreements rate is displayed in gray-scale blocks; the darker the block, the higher the disagreement rate. Figure 1 was created using Matlab R2017b (https://www.mathworks.com).

Figure 2

Examples showing the segmentation outcomes using the CNN-DS method. In each panel, the left, middle, and right images are the original CT slice, the ‘ground truth’ labels, and the CNN-DS predicted segmentation, respectively. The pointing arrows indicate the error. (A) Represents a case where the CNN-DS method demonstrates an expert-level performance. (B) Shows a false positive instance where a calcified structure is labelled as a hemorrhagic area due to its Hounsfield Unit values being higher than those of its surrounding tissues. (C) Shows a false negative example in which the CNN-DS method identified part of the hemorrhage but missed some blood close to the bone. (D) Illustrates a more complicated case of complex hemorrhage where the discrepancies between the ‘ground truth’ and the predicted segmentation cannot necessarily be attributed to erroneous prediction.

Figure 3

The training and validation loss for the U-Net model with and without deep supervision. The x-axis indicates the number of epochs, which is the number of times the deep learning model has passed through the entire training data during the training phase. The y-axis represents the loss value which implies how well the model behaves after each epoch; the lower the loss, the better a model. The dashed lines show the validation losses while the solid lines show the training losses. For the model with the deep supervision (blue lines), the training loss converges at a considerably faster rate, and the converged loss value is lower than the converged value of the model without deep supervision (green lines).

Figure 4

The architecture of the CNN-DS neural network model. The dashed lines show the skip connections while the solid lines show the normal ones. The neural network learns features of the image based on a hierarchy framework starting with simple features such as edges and shapes and going through more complex and high-level features in the deeper levels. The contracting path extracts the features while the expansive path reconstructs the final labelling. Google Slides was used to produce this figure (https://docs.google.com/presentation).