Deep Learning for Hemorrhagic Lesion Detection and Segmentation on Brain CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Stroke diagnosis, particularly hemorrhagic stroke, relies on accurate interpretation of computed tomographic (CT) images.
• Challenges in stroke diagnosis include high variability in lesion location, contrast, and shape, making manual detection time-consuming and difficult for radiologists.

Purpose of the Study:

• To develop and evaluate a U-net based deep learning framework for automated detection and segmentation of hemorrhage strokes in CT brain images.
• To enhance the accuracy and efficiency of stroke diagnosis using artificial intelligence.

Main Methods:

• A U-net based deep learning framework was designed, incorporating symmetry constraints by concatenating flipped and original CT slices.
• The framework explored various deep learning topologies, including modifications in layers, batch normalization, dilation rates, and pre-trained models.
• Adversarial training was employed to further improve segmentation accuracy.

Main Results:

• The proposed model achieved high performance on two datasets, with a location accuracy of 0.9859 for detection.
• Segmentation performance metrics included a Dice score of 0.8033 and an Intersection over Union (IoU) of 0.6919.
• The model demonstrated competitive performance compared to human experts.

Conclusions:

• The deep learning model effectively and robustly detects and segments hemorrhage strokes in CT images.
• The developed framework shows potential as a clinical decision support tool for improving stroke diagnosis.