# Machine learning-based radiomics for differentiating lung cancer subtypes in brain metastases using CE-T1WI

## TL;DR

This study uses machine learning and radiomic features from brain metastases imaging to distinguish between two types of lung cancer.

## Contribution

A novel machine learning-based radiomic approach using CE-T1WI for non-invasive differentiation of NSCLC and SCLC in brain metastases.

## Key findings

- 15 optimal radiomic features were selected from 833 extracted features for model training.
- LightGBM achieved the highest AUC of 0.853 in the test cohort for differentiating NSCLC and SCLC.
- Radiomic features from CE-T1WI effectively distinguish between NSCLC and SCLC in brain metastases.

## Abstract

The purpose of this research was to create and validate radiomic models based on machine learning that can effectively discriminate between primary non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC) in individuals with brain metastases (BMs) by utilizing high-dimensional radiomic characteristics derived from contrast-enhanced T1-weighted imaging (CE-T1WI).

A cohort of 260 individuals were chosen as participants. Among them, 173 individuals had NSCLC with 228 BMs, while 87 patients were diagnosed with SCLC with 142 BMs. Patients were allocated to a training dataset with a total of 259 BMs and an independent test dataset with a total of 111 BMs. Tumor tissues in axial CE-T1WI were manually outlined to delineate regions of interest (ROIs). Radiomic features were obtained from the ROIs using PyRadiomics, which were then chosen through a multistep selection process, including least absolute shrinkage and selection operator (LASSO) regression. Ten machine learning models, including Light Gradient Boosting Machine (LightGBM), RandomForest, and eXtreme Gradient Boosting (XGBoost), were built using selected features. The models’ performance was evaluated using receiver operating characteristic (ROC) analysis and area under the curve (AUC) calculations, complemented by additional metrics such as accuracy, specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV).

A total of 833 radiomic features were extracted from the ROIs. Through a multistep selection process, a refined subset of 15 optimal radiomic features was identified for model training. Ten classifier models were built based on features extracted from CE-T1WI. In the training dataset, the top-performing classifiers were the XGBoost, LightGBM, support vector machine (SVM) and random forest models, which achieved AUC of 0.963, 0.881, 0.876 and 0.855, respectively, with 5-fold cross-validation. Among the ten models tested, the LightGBM algorithm exhibited superior performance, with an AUC of 0.853 in the test cohort. This performance was superior to that of other models, such as RandomForest (AUC 0.843) and ExtraTrees (AUC 0.835). Radiomic features significantly contributed to the differentiation between NSCLC and SCLC.

Machine learning-based radiomics using CE-T1WI data is highly effective in distinguishing between NSCLC and SCLC in patients with BMs. The LightGBM model showed the best performance, suggesting that this approach shows promise as a supportive, non-invasive diagnostic tool, pending further validation in prospective clinical settings.