Deep Neural Network-Based Semantic Segmentation of Microvascular Decompression Images.

Area of Science:

• Medical Imaging
• Computer Vision
• Neuroscience

Background:

• Image semantic segmentation is crucial across various fields, including medicine.
• Accurate segmentation of cerebral vessels and cranial nerves from medical images remains a significant challenge.
• Existing methods require improvement for precise medical image analysis.

Purpose of the Study:

• To enhance the state-of-the-art DeepLabv3+ semantic segmentation network for medical imaging.
• To improve the accuracy and detail in segmenting cerebral vessels and cranial nerves.
• To refine feature extraction and boundary information retention in medical image segmentation.

Main Methods:

• Extended the DeepLabv3+ semantic segmentation network as the foundational framework.
• Introduced a feature distillation block (FDB) into the encoder for feature refinement.
• Integrated an atrous spatial pyramid pooling (ASPP) module into the decoder to preserve feature and boundary information.
• Fine-tuned and optimized model parameters for training.

Main Results:

• The enhanced encoder structure demonstrated superior performance in feature refinement.
• Significant improvements were observed in target boundary segmentation precision.
• The model successfully retained more critical feature information.
• Achieved a segmentation accuracy of 75.73%, outperforming the original DeepLabv3+ by 3%.

Conclusions:

• The proposed modifications to DeepLabv3+ effectively enhance medical image segmentation.
• The refined model shows significant potential for accurate cerebral vessel and cranial nerve segmentation.
• This advancement offers improved precision and information retention for medical image analysis.