DDANet: A deep dilated attention network for intracerebral haemorrhage segmentation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Neurological disorder diagnostics

Background:

• Intracranial haemorrhage (ICH) is a critical condition with high mortality, necessitating rapid and accurate detection.
• Deep neural networks show promise for ICH detection but struggle with small or subtle hemorrhages.
• Accurate segmentation of haematomas in brain CT images is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To introduce DDANet, a novel deep learning model for enhanced haematoma segmentation in brain CT images.
• To improve the detection accuracy of intracranial haemorrhage, particularly for subtle or small bleeds.
• To address the challenges of feature extraction and class imbalance in ICH segmentation.

Main Methods:

• Development of DDANet, a deep neural network incorporating a dilated convolution pooling block for enhanced feature extraction.
• Integration of a self-attention mechanism to capture global semantic information and guide feature processing.
• Incorporation of residual networks, channel attention, and spatial attention mechanisms for joint optimization and class imbalance mitigation.

Main Results:

• DDANet achieved superior performance compared to existing methods in haematoma segmentation.
• The model obtained a Dice coefficient of 0.712, Jaccard index of 0.601, sensitivity of 0.73, accuracy of 0.997, and specificity of 0.998.
• Experimental results demonstrate DDANet's effectiveness in handling subtle hemorrhages and class imbalance.

Conclusions:

• DDANet represents a significant advancement in automated intracranial haemorrhage segmentation on brain CT.
• The model's architecture effectively enhances feature extraction, global context understanding, and robustness to class imbalance.
• DDANet shows potential for improving the speed and accuracy of ICH diagnosis in clinical settings.