Pathobiological Dictionary Defining Pathomics and Texture Features: Addressing Understandable AI Issues in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Artificial intelligence (AI) shows promise in medical diagnostics, but its clinical adoption is hindered by a lack of interpretability.
• Translating complex imaging features into clinically meaningful insights is crucial for integrating AI into diagnostic workflows.

Purpose of the Study:

• To develop and validate the Pathobiological Dictionary for Liver Cancer (LCP1.0), a framework to bridge AI-derived Pathomics (PF) and Radiomics Features (RF) with clinical semantics.
• To enhance the transparency and usability of AI models in liver cancer diagnostics.

Main Methods:

• Extracted 333 imaging features (PF and RF) from hepatocellular carcinoma (HCC) tissue samples using QuPath and PyRadiomics, adhering to IBSI guidelines.
• Utilized expert-defined regions of interest and feature selection algorithms (Variable Threshold with SVM) to identify key features for survival outcome prediction.
• Validated the LCP1.0 dictionary with 8 oncology and pathology experts.

Main Results:

• The Variable Threshold feature selection with SVM achieved a cross-validation accuracy of 0.80 ± 0.01, selecting 20 key features.
• Identified features related to cell nucleus and cytoplasm characteristics (e.g., Centroid, Cell Nucleus, Cytoplasmic) as strongly associated with tumor grading and prognosis.
• The selected features reflect indicators of atypia, such as pleomorphism, hyperchromasia, and cellular orientation.

Conclusions:

• The LCP1.0 provides a clinically validated bridge between AI outputs and expert interpretation for liver cancer pathology.
• Aligning AI-derived features with clinical semantics supports the development of interpretable and trustworthy diagnostic tools.
• This framework facilitates the adoption of AI in clinical practice by enhancing model transparency and usability.