Combating Ambiguity for Hash-Code Learning in Medical Instance Retrieval.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical instance retrieval aids radiologists in diagnosis by finding similar cases.
• Pathological region features often lack specificity, leading to diagnostic ambiguity.
• Current methods struggle with varying disease manifestations and stages.

Purpose of the Study:

• To develop a novel deep framework, Y-Net, to address manifestation ambiguity in medical instance retrieval.
• To improve the accuracy and discriminative power of medical image retrieval systems.

Main Methods:

• Y-Net encodes images into compact hash-codes using aggregated convolutional features.
• It unifies pixel-wise segmentation loss and classification loss to learn discriminative features.
• Segmentation loss enhances spatial discriminability, while classification loss improves semantic separability.

Main Results:

• Y-Net effectively enhances visual features in abnormal regions and suppresses background noise.
• The framework alleviates ambiguity associated with pathological region manifestations.
• Achieved an average performance improvement of 9.27% over state-of-the-art methods in retrieval tasks.

Conclusions:

• Y-Net offers a robust solution for enhancing medical instance retrieval by tackling manifestation ambiguity.
• The proposed deep learning approach significantly improves retrieval performance and diagnostic support.