More than just a heatmap: elevating XAI with rigorous evaluation metrics.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Neuroscience

Background:

• Deep learning (DL) models show promise in tumor diagnosis and treatment planning using MRI and ultrasound.
• However, high computational demands and limited explainability of DL hinder clinical adoption.
• Existing explainable AI (XAI) methods often produce fragmented or misaligned saliency maps.

Purpose of the Study:

• To develop SpikeNet, a hybrid framework combining CNNs and SNNs for efficient and explainable tumor analysis.
• To introduce XAlign, a metric for quantifying the alignment of AI-generated explanations with expert annotations.
• To evaluate SpikeNet's performance and explainability against traditional XAI methods.

Main Methods:

• SpikeNet integrates CNNs for spatial encoding and SNNs for event-driven processing, featuring a native saliency module for real-time explanations.
• The XAlign metric assesses explanation fidelity by considering regional concentration, boundary adherence, and dispersion.
• Cross-validation was performed on TCGA-LGG (MRI) and BUSI (ultrasound) datasets, aggregating slice-level predictions to patient-level decisions.

Main Results:

• SpikeNet achieved high accuracy and F1 scores on both TCGA-LGG (97.12% accuracy, 97.43% F1) and BUSI (98.23% accuracy, 98.32% F1) datasets.
• The framework demonstrated low per-image latency (approx. 31ms) and high throughput (approx. 32 images/sec) on a single GPU.
• SpikeNet's explanations, evaluated by XAlign, showed superior alignment compared to Grad-CAM, LIME, and SHAP.

Conclusions:

• SpikeNet provides accurate, low-latency, and explainable tumor analysis for MRI and ultrasound, enhancing clinical decision support.
• The XAlign metric offers a clinically relevant assessment of explanation fidelity, enabling robust comparisons between methods.
• SpikeNet and XAlign show significant potential for trustworthy and efficient clinical AI applications.