Boosting Few-Shot Semantic Segmentation of 3D Medical Images via Collaborative Slice Alignment.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot semantic segmentation (FSS) in 3D medical imaging relies on selecting appropriate support slices from labeled volumes to segment unlabeled query slices.
• Accurate support slice selection is vital for learning robust prototypical features and improving segmentation accuracy.
• Current methods often use true target locations or single support slices, limiting practicality and leading to segmentation errors.

Purpose of the Study:

• To propose a practical and efficient solution for few-shot semantic segmentation (FSS) of 3D medical images.
• To develop a novel Collaborative Slice Alignment (CSA) module for improved support slice selection.
• To enhance segmentation accuracy by enabling each query slice to find its fittest support slice without prior target information.

Main Methods:

• Proposed a Collaborative Slice Alignment (CSA) module that estimates slice confidence scores to reflect physical location, enabling spatial alignment of support and query slices.
• Implemented a self-learnable ranking objective within CSA to transfer internal knowledge between support and query features, boosting FSS performance.
• Introduced an Information Reconciliation (InRe) module to address inconsistent feature distributions between support and query images.

Main Results:

• The combined CSA and InRe modules achieved a significant average Dice score improvement of at least 8.61% across three diverse datasets.
• The proposed method consistently outperformed existing state-of-the-art methods in 3D medical image segmentation.
• CSA effectively aligned support and query slices based on estimated spatial relevance, leading to more accurate segmentation.

Conclusions:

• The novel CSA module offers a practical and efficient approach to few-shot semantic segmentation in 3D medical imaging.
• The InRe module effectively mitigates feature distribution inconsistencies, further enhancing segmentation performance.
• The proposed method demonstrates superior accuracy and robustness compared to current state-of-the-art techniques.