Saliency-driven explainable deep learning in medical imaging: bridging visual explainability and statistical quantitative analysis | JoVE Visualize

Area of Science:

Medical Image Analysis
Artificial Intelligence in Healthcare
Explainable AI (XAI)

Background:

Deep learning models show promise in medical imaging but lack explainability, hindering clinical trust and adoption.
Current research often relies on visual inspection rather than quantitative analysis for attribution in medical imaging.
Explainability techniques are crucial for increasing stakeholder confidence in AI-driven healthcare solutions.

Purpose of the Study:

To propose and evaluate an image-based saliency framework for enhancing the explainability of deep learning models in medical image analysis.
To integrate qualitative and statistical quantitative assessments for a comprehensive evaluation of saliency methods.
To improve the transparency and trustworthiness of deep learning models in clinical settings.

Main Methods:

Utilized adaptive path-based gradient integration, gradient-free techniques, and class activation mapping (CAM) derivatives.
Applied the framework to brain tumor MRI and COVID-19 chest X-ray datasets using deep convolutional neural networks.
Employed Accuracy Information Curves (AICs) and Softmax Information Curves (SICs) for quantitative assessment of saliency effectiveness.

Main Results:

Visual inspection showed ScoreCAM, XRAI, GradCAM, and GradCAM++ produced clinically interpretable attribution maps, highlighting biomarkers and model biases.
Empirical evaluations indicated ScoreCAM and XRAI were effective in retaining relevant image regions, evidenced by higher AUC values.
Softmax Information Curves (SICs) revealed variability, with some random saliency masks outperforming established methods, stressing the need for combined evaluation metrics.

Conclusions:

Combining qualitative and quantitative approaches is essential for comprehensive evaluation and enhances the transparency and trustworthiness of deep learning models.
Selecting appropriate saliency methods tailored to specific medical imaging tasks is crucial for effective model explainability.
This work advances model explainability to foster greater trust and clinical adoption of AI in healthcare, with future work focusing on refining metrics and expanding modalities.