Adaptive Multi-Prototype Open-Set Recognition Network for Medical Image Classification

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Open-set recognition (OSR) methods often use fixed prototypes, failing to capture intra-class variations common in medical images.
• High intra-class heterogeneity in medical data presents a significant challenge for accurate OSR.

Purpose of the Study:

• To propose a novel framework, the Adaptive Multi-Prototype Network with Pretrained Swin Transformer (PSW-AMPN), for improved OSR in medical imaging.
• To address the limitations of fixed prototype methods in handling complex intra-class structures.

Main Methods:

• Developed a sparse gated attention module to adaptively select discriminative sub-prototypes based on prototype-sample relations.
• Employed a Pretrained Swin Transformer and a lightweight projector for effective feature extraction, capturing both local and global image features.
• Jointly optimized classification loss with regularization for open space risk using multiple prototypes.

Main Results:

• The proposed PSW-AMPN framework effectively captures complex intra-class structures and enhances class discriminability.
• Achieved state-of-the-art performance on multiple medical image OSR tasks, significantly outperforming existing baseline methods.
• Demonstrated the capability to handle high intra-class heterogeneity inherent in medical imaging data.

Conclusions:

• The PSW-AMPN framework offers a robust solution for open-set recognition in medical imaging by adaptively managing prototypes.
• The approach enhances the accuracy and reliability of OSR in medical applications with significant intra-class variations.
• This work advances the field of medical image analysis by providing a more effective tool for distinguishing known from unknown classes.