XRelevanceCAM: towards explainable tissue characterization with improved localisation of pathological structures in...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurosurgery

Background:

• Probe-based confocal laser endomicroscopy (pCLE) aids intraoperative brain tumor characterization, potentially improving resection rates.
• Deep learning models offer automated tissue analysis but often lack transparency, hindering clinical trust and interpretability.
• Existing explanation methods like Class Activation Map (CAM) variants struggle with transparency and the shattered gradient problem.

Purpose of the Study:

• To introduce XRelevanceCAM, a novel explanation method for deep learning models used in pCLE.
• To provide human-interpretable visual explanations for surgical decision support in brain tumor resection.
• To address the limitations of existing CAM techniques, including transparency and the shattered gradient problem.

Main Methods:

• Developed XRelevanceCAM based on an improved backpropagation approach incorporating sensitivity and conservation axioms.
• Utilized ex vivo pCLE data from brain tumors for qualitative and quantitative evaluations.
• Compared XRelevanceCAM against baseline methods, including RelevanceCAM, using mean Intersection over Union (mIoU) with ground-truth annotations.

Main Results:

• XRelevanceCAM effectively highlights clinically relevant tissue areas in pCLE images.
• Achieved a 56% improvement in mIoU in the shallowest network layer compared to the closest baseline (RelevanceCAM).
• Demonstrated a 6% improvement in mIoU when generating saliency maps from all network layers.

Conclusions:

• XRelevanceCAM is a new Class Activation Map (CAM) variation for precise identification of critical structures in pCLE data.
• The method offers enhanced transparency and interpretability for deep learning models in surgical contexts.
• XRelevanceCAM can significantly aid intraoperative decision support during brain tumor resection surgeries.