# Development and validation of multiple machine learning algorithms for differentiating primary central nervous system lymphoma from adult-type diffuse glioma: an interpretable and multicenter study

## TL;DR

This study develops an interpretable machine learning model using MRI and clinical data to distinguish between two brain tumors, improving preoperative decision-making for neurosurgeons.

## Contribution

A novel interpretable radiomic-clinical fusion model using SHAP values and a nomogram for accurate differentiation of PCNSL and ADG.

## Key findings

- The Rad-Clinic fusion model achieved an AUC of 0.973 in training and 0.940 in external validation.
- CET1WI+DWI+FLAIR fusion model had an AUC of 0.871 in external validation.
- SHAP values provided interpretable insights into feature contributions for model predictions.

## Abstract

Preoperative differentiation of primary central nervous system lymphoma (PCNSL) from adult-type diffuse glioma(ADG) is important to guide neurosurgical decision-making.To develop and validate a MRI–based interpretable radiomic-clinical(Rad-Clinic) fusion model to differentiate PCNSL from ADG by seven machine learning algorithms.

In this retrospective study, we recruited 165 patients who underwent preoperative conventional MRI(CET1WI, FLAIR, DWI, ADC) with PCNSL and ADG from two institutions (115 in the training cohort and 50 in the external validation cohort). we selected seven machine learning algorithms to construct a framework incorporating radiomic features and clinical parameters. SHapley Additive exPlanations (SHAP) values elucidated feature contributions, and a radiomic nomogram was developed for clinical translation.

The CET1WI+DWI+FLAIR fusion model exhibited optimal performance among all the single-sequence and multi-sequence radiomic models, and the AUC for external validation cohort were 0.871. But the Rad-Clinic fusion model performed well in differentiating PCNSL from ADG, and the AUC for the training and external validation cohort were 0.973 and 0.940, outperforming radiomic model and clinical model.SHAP summary plot illustrated the feature’s value affected the feature’simpact attributed to the Rad-Clinic fusion model.The nomogram demonstrated clinical interpretability through visualised risk stratification.

An interpretable Rad-Clinic fusion model enables accurate preoperative to differentiate PCNSL from ADG, and may assist improve clinical decision-making.