Acute ischemic stroke lesion segmentation in non-contrast CT images using 3D convolutional neural networks

TL;DR

This paper presents a deep learning-based method using a modified 3D U-Net architecture with advanced pre-processing and sampling strategies for automatic segmentation of acute ischemic stroke lesions in non-contrast CT images, achieving promising results.

Contribution

The study introduces a novel 3D CNN model with squeeze-and-excitation and residual blocks, along with specialized pre-processing and sampling, for improved stroke lesion segmentation in CT scans.

Findings

Achieved an average Dice coefficient of 0.628

Sensitivity of 0.699 indicates good lesion detection

High specificity of 0.9965 reduces false positives

Abstract

In this paper, an automatic algorithm aimed at volumetric segmentation of acute ischemic stroke lesion in non-contrast computed tomography brain 3D images is proposed. Our deep-learning approach is based on the popular 3D U-Net convolutional neural network architecture, which was modified by adding the squeeze-and-excitation blocks and residual connections. Robust pre-processing methods were implemented to improve the segmentation accuracy. Moreover, a specific patches sampling strategy was used to address the large size of medical images, to smooth out the effect of the class imbalance problem and to stabilize neural network training. All experiments were performed using five-fold cross-validation on the dataset containing non-contrast computed tomography volumetric brain scans of 81 patients diagnosed with acute ischemic stroke. Two radiology experts manually segmented images independently and then verified the labeling results for inconsistencies. The quantitative results of the proposed algorithm and obtained segmentation were measured by the Dice similarity coefficient, sensitivity, specificity and precision metrics. Our proposed model achieves an average Dice of $0.628\pm0.033$, sensitivity of $0.699\pm0.039$, specificity of $0.9965\pm0.0016$ and precision of $0.619\pm0.036$, showing promising segmentation results.