Quantification-based explainable artificial intelligence for deep learning decisions: clustering and visualization of
Gen Takagi1, Saori Takeyama1, Tokiya Abe2
1Institute of Science Tokyo, School of Engineering, Department of Information and Communications Engineering, Yokohama, Japan.
Journal of Medical Imaging (Bellingham, Wash.)
|October 13, 2025
Summary
This study introduces a quantitative explainable artificial intelligence (QXAI) method to interpret deep learning (DL) decisions in liver cancer pathology. The approach enhances trust and clinical adoption of DL tools by revealing key morphological features driving diagnoses.
Area of Science:
- Computational Pathology
- Artificial Intelligence in Medicine
- Digital Pathology
Background:
- Deep learning (DL) models achieve high accuracy in computational pathology but often lack transparency, hindering clinical integration.
- The "black box" nature of DL necessitates the development of interpretable AI methods for reliable medical applications.
- Quantitative Explainable AI (QXAI) is crucial for understanding DL decision-making processes in complex diagnostic tasks.
Purpose of the Study:
- To develop and validate a QXAI approach for objectively interpreting DL model decisions in hepatocellular carcinoma (HCC) pathological image analysis.
- To quantitatively elucidate the reasoning behind DL classifications by identifying critical image regions and morphological features.
- To enhance the clinical adoption of DL in pathology by providing transparent and trustworthy diagnostic insights.
Main Methods:
- Utilized latent space clustering of DL embeddings to identify discriminatory image regions.
- Quantitatively characterized identified regions using morphometric features derived from nuclear segmentation (HoverNet).
- Employed LightGBM for key feature selection and statistical analysis to link morphology to classification outcomes.
Main Results:
- The QXAI method successfully identified key discriminatory regions and features in HCC pathology images.
- Morphological features such as nuclear size, chromatin density, and shape irregularity were found to be critical.
- Clustering-based analysis provided structured, clinically relevant, and pathologist-validated insights into classification drivers.
Conclusions:
- The proposed QXAI framework effectively links morphological features to DL classification decisions in HCC analysis.
- This approach significantly enhances the interpretability and trustworthiness of DL models in digital pathology.
- QXAI facilitates the seamless clinical integration of advanced AI tools for improved cancer diagnosis.
