Coarse for Fine: Bounding Box Supervised Thyroid Ultrasound Image Segmentation Using Spatial Arrangement and...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Weakly-supervised learning is attractive for medical image segmentation but struggles with precise boundary delineation in thyroid ultrasound images due to corrupted low-level features.
• Existing methods often overfit to weakly annotated regions, failing to accurately segment targets.

Purpose of the Study:

• To propose a novel dual-branch weakly-supervised learning framework for optimizing medical image segmentation.
• To enhance the spatial distribution calibration of semantic features using coarse bounding box guidance.
• To improve the accuracy of thyroid gland and nodule segmentation.

Main Methods:

• A dual-branch framework calibrates semantic features for rational spatial distribution under bounding box mask guidance.
• The spatial arrangement consistency branch compares maximum activations to measure target localization rationality.
• The hierarchical prediction consistency branch uses prototypes to quantify semantic feature perception rationality.

Main Results:

• The proposed model outperforms existing weakly-supervised methods in thyroid gland and nodule segmentation.
• Performance is comparable to fully-supervised methods, significantly reducing annotation time.
• The method effectively considers target location and semantic feature distribution rationality.

Conclusions:

• The dual-branch framework offers a robust weakly-supervised segmentation strategy for medical imaging.
• This approach enhances the clinical applicability of deep learning-based segmentation for thyroid ultrasound.
• It provides precise delineation of target boundaries by addressing limitations of existing methods.