DSIFNet: Implicit feature network for nasal cavity and vestibule segmentation from 3D head CT

Area of Science:

• Medical Imaging
• Computer Vision
• Anatomy

Background:

• Accurate segmentation of the nasal cavity and its internal structures from head CT images is crucial for clinical applications.
• Nasal anatomy presents challenges due to significant scale differences, complex shapes, and variable microstructures, requiring advanced segmentation techniques.
• Existing methods may struggle with cross-scale feature extraction essential for detailed nasal structure analysis.

Purpose of the Study:

• To develop and validate a novel deep learning network for accurate segmentation of the nasal cavity and vestibule in head CT images.
• To enhance the network's ability to handle cross-scale features and improve segmentation precision for intricate anatomical details.
• To leverage large-scale datasets and self-supervised pretraining for robust feature extraction in medical image segmentation.

Main Methods:

• Proposed the Deeply Supervised Implicit Feature Network (DSIFNet) incorporating an Implicit Feature Function Module Guided by Local and Global Positional Information (LGPI-IFF).
• Implemented a deep supervision mechanism utilizing implicit feature functions in the decoding phase to optimize multi-scale feature utilization.
• Constructed a large dataset of 7116 CT volumes and employed PixPro-based self-supervised pretraining on unlabeled data.

Main Results:

• The DSIFNet demonstrated robust and superior performance in nasal cavity and vestibule segmentation on a test set of 128 head CT volumes.
• Achieved leading results across multiple segmentation metrics, indicating high precision and improved detail representation.
• The LGPI-IFF module effectively fused features across scales, enhancing the recognition of both fine details and overall structures.

Conclusions:

• The proposed DSIFNet effectively addresses the challenges of cross-scale feature extraction for nasal cavity segmentation in CT images.
• The deep supervision mechanism and specialized modules contribute to improved segmentation accuracy and detailed anatomical representation.
• The method shows significant potential for advancing nasal physiology studies, disease diagnosis, and surgical planning through enhanced medical image analysis.