Deep learning for liver tumor diagnosis part II: convolutional neural network interpretation using radiologic imaging...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models excel at image classification but often lack transparency.
• Interpretable AI is crucial for clinical adoption, especially in medical diagnostics.
• Hepatic lesion classification on MRI requires expert analysis of subtle imaging features.

Purpose of the Study:

• To develop a proof-of-concept interpretable deep learning system for hepatic lesion classification.
• To enable the AI model to justify its predictions by identifying key imaging features.
• To enhance the clinical utility of AI in radiology through explainability.

Main Methods:

• A convolutional neural network (CNN) was trained to classify six hepatic tumor entities using multi-phasic MRI data.
• A post hoc algorithm analyzed CNN activation patterns to infer and map key imaging features.
• Feature relevance scores were assigned to quantify the contribution of each feature to the classification.

Main Results:

• The interpretable system achieved 76.5% positive predictive value and 82.9% sensitivity in identifying radiological features.
• Feature maps accurately highlighted image regions corresponding to identified features.
• Feature relevance scores generally aligned with established radiological criteria for lesion classification.

Conclusions:

• The developed system demonstrates the feasibility of interpretable deep learning in medical imaging.
• This approach can illuminate AI decision-making processes, identifying and localizing critical imaging features.
• Interpretable AI has the potential to support radiologists in differential diagnosis and reporting, enhancing clinical practicality.