Lymphovascular Invasion prediction in Non-Small Cell Lung Carcinoma Using Multimodality PET/CT Fusion Radiomics

Introduction: Lymphovascular invasion (LVI) is one of the main factors of poor prognosis in Non-Small Cell Lung Carcinoma (NSCLC) patients since it is often followed by local recurrence, and distant metastasis of the tumor. However, LVI prediction is challenging, considering that it is a histopathologic condition that is currently diagnosed from specimens obtained by biopsy, which is an invasive and expensive procedure. In this study, toward identifying optimum performance, we proposed a framework to investigate the prognostic power of single- and multi-modality PET/CT fused radiomics models, trained by various feature selection (FS) and machine learning (ML) methods for prediction of LVI in NSCLC patients.

Methods: We retrospectively enrolled 146 NSCLC (126 LVI positive, 20 LVI negative) patients which underwent PET/CT scan prior to surgery. Tumors were delineated on CT scans by an experienced radiologist. Intensity range of CT scans were clipped to the lung window and then normalized, while PET SUV’s were normalized to the median of the whole dataset. Prior to fusion, images were registered and interpolated to isotropic voxel spacing of 1mm³. For image fusion two different methods including Guided Filtering Fusion (GFF), and one based on visual saliency map and Weighted Least Square (WLS) were utilized. Morphological, intensity based, and textural radiomic features were extracted using SERA package, which follows IBSI guidelines. Combination of three feature selection methods namely, Boruta, Maximum Relevance Minimum Redundancy (MRMR), and Recursive Feature Elimination (RFE), and eight machine learning algorithms namely Decision Tree (DT), K-Nearest neighbor (KNN), Logistic Regression (LR), Multi-Layer Perceptron (MLP), Naive Bayes (NB), Random Forest (RF), Support Vector Machine (SVP), and regularized gradient boosting (XGB) were applied to the radiomics feature-sets. Dataset was divided to 70-30% partitions for training/validating and testing the models. The trained model was applied 1000 times to the test cohort using bootstrapping and reported results were averaged to repel possible biases. Grid search was utilized for hyperparameter optimization. Performance of the models were reported with area under ROC curve (AUC), accuracy (ACC), Sensitivity (SEN), and Specificity (SPE).

Results: For CT, combination of RFE as FS and KNN as ML algorithms yielded highest performance (AUC, ACC, SEN, SPE = 0.81,0.72, 0.84, 0.71, respectively). Regarding PET models, model based on Boruta as FS and KNN as ML achieved best performance (0.80, 0.75, 0.83, 0.73). Multimodality models fused by GFF, achieved the highest results on average, which among them model trained by MRMR as FS and LR as ML methods reached best outcome (0.81, 0.70, 0.84, 0.68).

Conclusions: NSCLC Lymphovascular invasion can be predicted non-invasively by single- and multimodality PET/CT radiomics modeling. Radiomic features extracted from multimodality fused images can better characterize tumor heterogeneities, since they simultaneously reflect both anatomical (CT), and metabolic (PET) aspects of the tumor. Also, there was no fit to all algorithm, and different modalities achieved their best performance while trained by different combination of FS and ML algorithms.

• Download figure
• Open in new tab
• Download powerpoint