Interpretable classification of endomicroscopic brain data via saliency consistent contrastive learning.

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate brain tissue characterization using probe-based Confocal Laser Endomicroscopy (pCLE) is crucial for surgical guidance and tumor resection.
• Current deep learning models for pCLE data lack interpretability, hindering surgeon trust.
• Existing classification methods often use cross-entropy loss, which may not optimize for distinguishing tissue classes effectively.

Purpose of the Study:

• To develop a novel, interpretable image classification framework for brain tissue characterization using pCLE data.
• To improve the accuracy, robustness, and interpretability of AI models in neurosurgery.
• To provide surgeons with reliable AI tools that offer clear insights into tissue classification.

Main Methods:

• Proposed a Label Contrastive Learning (LCL) loss to enhance intra-category similarity and inter-category contrast, generating representative data embeddings.
• Introduced a Saliency Consistency (SC) module with a Top-K Maximum and Minimum Pooling (TK-MMP) layer for clinically relevant saliency map generation.
• Utilized Exponential Moving Average (EMA) for global tissue category embeddings and a Global Embedding Inference (GEI) layer for robust classification via cosine similarity.

Main Results:

• The proposed framework achieved superior classification performance compared to state-of-the-art models on both ex-vivo and in-vivo pCLE brain data.
• Demonstrated significant improvements in accuracy, robustness, and interpretability of brain tissue classification.
• Generated clinically relevant saliency maps, enhancing the understanding of model predictions.

Conclusions:

• The novel LCL and SC framework provides a highly accurate, robust, and interpretable solution for brain tissue characterization using pCLE.
• This approach has the potential to significantly enhance surgeon trust and decision-making in neurosurgical procedures.
• The developed methods offer a new standard for interpretable AI in medical imaging applications.