Lcp1.0: A Computational Pathomics Dictionary As a Translational Tool Between Ai Researchers and Clinicians in Liver Cancer Care Publisher



Salmanpour MR ; Piri SM ; Mehrnia SS ; Rahmim A ; Hacihaliloglu I
Source: Progress in Biomedical Optics and Imaging - Proceedings of SPIE Published:2026

Abstract

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality, with survival dependent on early diagnosis and accurate tumor grading. Histopathology is the gold standard, yet time-consuming, prone to inter-observer variability, and limited in capturing tumor heterogeneity. Computational pathology and pathomics (PF)/radiomics (RF) extract quantitative features from tissue slides and imaging, but these remain opaque to clinicians, creating a gap between AI insights and actionable knowledge. This slows AI translation despite its potential to improve diagnostics and treatment planning. To address this, we present the Liver Cancer Pathobiological Dictionary (LCP1.0), a collaborative framework translating PFs and RFs into actionable insights. By mapping AI-derived metrics to recognizable traits, LCP1.0 bridges computational outputs and clinical decision-making, enhancing trust and interpretability of AI in liver cancer diagnostics. We analyzed HCC tissue samples using QuPath and PyRadiomics (IBSI-compliant), extracting 333 imaging features: 240 PF-based cellular/intensity metrics, 74 texture-RFs, and 19 first-order RFs. Expert-curated 1000×1000 µm regions of interest (ROIs) excluded artifacts, with features aggregated per case. Feature relevance to WHO grading was assessed using classifier-feature selector pipelines with three feature selection algorithms and five classification algorithms, serving as an example of the dictionary's application. This approach can be expanded to other predictive tasks, enabling broader use of the dictionary in various clinical settings. The dictionary was validated by eight oncologists/pathologists. Collaborating with ten specialists, we mapped PF/RF features to pathobiological traits. The SVM model with Variable Threshold selection achieved optimal accuracy (0.80±0.05, p<0.05), identifying 20 key features—including nuclear (Centroid, hyperchromasia) and cytoplasmic metrics—strongly correlated with tumor grade and survival. These features reflect histopathological atypia (pleomorphism, cellular disorganization), demonstrating their diagnostic relevance. LCP1.0 bridges AI-derived data and clinical intuition by anchoring computational features to established pathological semantics. This framework enhances interpretability, fostering trust in AI-assisted liver cancer diagnostics while maintaining alignment with routine practice. © 2026 COPYRIGHT SPIE.