# Interpretable ADC-based radiomics models for differentiating hepatocellular carcinoma and intrahepatic cholangiocarcinoma

## TL;DR

This study developed a machine learning model using MRI data to distinguish between two types of liver cancer, showing strong performance and potential for non-invasive diagnosis.

## Contribution

The novel contribution is an interpretable logistic regression model using ADC radiomics with validated generalizability for differentiating HCC and ICC.

## Key findings

- The LR model achieved high AUROCs (0.95 in training, 0.91 in internal validation, 0.85 in external validation).
- Calibration plots and decision curve analysis confirmed the model's clinical utility and robustness.
- Wavelet-LLL-firstorder-RootMeanSquared was identified as the most impactful feature via SHAP analysis.

## Abstract

This study aimed to develop interpretable machine learning (ML) models using apparent diffusion coefficient (ADC) radiomics to differentiate hepatocellular carcinoma (HCC) from intrahepatic cholangiocarcinoma (ICC).

Radiomic features were extracted from ADC maps of 83 pathologically confirmed HCC and 46 pathologically confirmed ICC patients who underwent MRI examinations. The least absolute shrinkage and selection operator (LASSO) method selected essential features for five ML models: logistic regression (LR), random forest (RF), gaussian naive bayes (GNB), support vector machine (SVM), and k-nearest neighbors (kNN). external validation was performed using 20 HCC and 20 ICC cases from the cancer imaging archive (TCIA) public database. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, accuracy, F1 score, calibration plots, and decision curve analysis (DCA). The best-performing model was interpreted using shapley additive explanations (SHAP).

LASSO selected eight features. The models achieved training AUROCs of 0.84-0.95 and internal validation AUROCs of 0.78-0.91. The LR model demonstrated superior performance (training AUROC: 0.95; internal validation AUROC: 0.91; external validation AUROC: 0.85). Moreover, calibration plots and DCA confirmed that this model exhibited the best calibration and clinical utility. SHAP identified wavelet-LLL-firstorder-RootMeanSquared as the most impactful feature.

The ADC-based LR model robustly differentiates HCC from ICC, with validated generalizability using public data, offering a promising non-invasive clinical tool.