Machine Learning based Classifier for lytic bone lesions of Langerhans cell histocytosis (LCH) and multiple myeloma (MM).

Introduction: Langerhans cell histocytosis (LCH) and multiple myeloma (MM) commonly present with lytic bone lesions. These lesions can be differentiated based on clinical history and histopathology. However, radiologically, there is significant overlap in the finding of these two conditions on F18-FDG-PET/CT. The present study was aimed to find out role of Machine Learning based Classifier to differentiate lytic bone lesions of LCH and MM.

Methods: The trans-axial F18-FDG-PET/CT image of histopathologically proven patients of LCH (n=77) and Multiple Myeloma (n=85). The lesion was segmented and texture of the lesion {26 Haralick texture features (13 at 1 pixel, and another 13 at 2-pixel distance)} were computed. Oversampling method was used to balance the dataset and then features having correlation greater than 0.70 were dropped. The remaining four feature vectors were used for training and testing the machine learning model (SVMradial) for classifying the LCH and Multiple myeloma images. The following procedures were repeated fifty times: "random partition of dataset into training and test dataset in the ratio of 70:30, 10-fold cross validation techniques during training, testing the performance of the trained model on test dataset". The model having best performance was selected. All experiments were performed on personal computer in R using the "ROSE" and "CARET" package.

Results: Of 77 LCH patients (mean age = 7.89 ± 9.76 years; median age = 4 years, mode = 2 years, range = 0.5 years to 47 years), 25 patients were female and 52 patients were male. Of 85 Multiple Myeloma patients (mean age = 51.83 ± 9.65 years; range: 31 years to 81 years) 50 patients were male and 35 patients were female. The best performing model (SVMradial) was found to have 89% accuracy, 92% sensitivity and 85% specificity on training dataset. When machine learning SVMradial model was applied on test data set in classifying the lytic lesions of LCH and multiple Myeloma, it was found to have accuracy of 80%, sensitivity of 72%, and specificity of 88%.

Conclusions: The trained machine learning model, SVMradial classifier, was found to differentiate lytic bone lesions of Langerhans cell histocytosis (LCH) and Multiple myeloma with high specificity and accuracy.