Radiomics and Machine Learning in Gall Bladder cancer vs cholecystitis: simplifying the diagnostic dilemma

Introduction: Gall bladder cancer (GBCa) is an aggressive malignancy with poor prognosis. However, its diagnosis can be challenging and is often confused with cholecystitis, particularly xanthogranulomatous cholecystitis. Neither clinical symptoms, laboratory tests, nor conventional radiological techniques offer a reliable method for differential diagnosis. Thus, we investigated the utility of 18F-FDG PET-CT derived radiomic features (RF’s) and a machine learning model (MLM) to distinguish GBCa from cholecystitis accurately. Materials and Methods: Patients with suspected GBCa undergoing 18F-FDG PET-CECT were included. Experienced nuclear medicine physicians manually delineated the regions of interest (ROIs), and intensity-intensity-threshold restrictions of 25% and 35% were also applied to create three distinct sets of ROIs and high-dimensional RF’s were extracted using LIFEx (v7.8). Robust RFs were filtered using univariate analysis and Mann-Whitney U test, preserving those with AUC >0.7 and p value <0.05. Histopathological examination (HPE) was taken as the gold standard. Selected RF’s were further refined using a total of 5 feature selection methods: random forest (RF), distance correlation (DC), eXtreme gradient boosting (Xgboost), gradient boosting decision tree (GBDT) and least absolute shrinkage and selection operator (LASSO). Nine MLMs were applied: support vector machine (SVM), linear discriminant analysis (LDA), decision tree (DT), Random Forest, K-nearest neighbourhood (KNN), GBDT, adaptive boosting (AdaBoost), logistic regression (LR) and Gaussian Naïve-Bayes (GaussianNB) were developed using Python-3 with 75% of data for training and 25% for validation. The model's performance was assessed using ROC curves and compared with metabolic parameters. Similarly, another MLM was developed to predict the probability of metastatic disease. Results: The study included 41 GBCa patients (14 males, median age-57, range: 35-76 years). Among these, 35 had GBCa while and 6 had cholecystitis on final HPE. A total of 161- RF’s were extracted from each 18-F FDG PET-CT image at three intensity thresholds, with the most significant results observed at 25% threshold. Among MLM’s, Gaussian NB MLM created using features selected by Distance Correlation feature selection methods out performing all other MLM’s with an accuracy of 63.6%, sensitivity of 55.6%, specificity of 100% with an AUC of 0.944. While for metastatic prediction GBDT with features selected using Random Forest at 25% threshold levels demonstrating most robustness achieving a sensitivity of 100%, specificity of 75%, and accuracy of 90%, with an AUC of 0.944 Conclusion: RF’s extracted from functional imaging modalities like 18-F FGDG PET-CT and a machine learning model (MLM) can be instrumental in effectively predicting the histology (GBCa vs cholecystitis) as well as distant metastasis in cases with Suspected GBCa.