Abstract
P595
Introduction: LCH and Tuberculosis commonly present 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 evaluate the use of machine learning on the Haralick texture feature of the lytic lesion with random forest algorithm as a classifier for lytic bone lesions of LCH and Tuberculosis.
Methods: The study included trans-axial F18-FDG-PET/CT images of histopathologically proven patients of LCH (n=77) and Tuberculosis patients (n=42). The lytic lesions were segmented and the gray level co-occurrence matrix (GLCM) of the segmented image was created with 32 gray levels and 26 Haralick texture features (13 at 1-pixel, and the other 13 at 2-pixel distance) were estimated. Texture features that had correlation less than 0.70 were used for training and testing the random forest model. The datasets were balanced using oversampling technique. Fifty different models were trained and tested. Different training and test datasets for use in each model were generated by random partition of the dataset into 80:20 ratio. Ten-fold cross validation was used during training. The model having highest accuracy, sensitivity and specificity was selected. The experiment was performed in R programming language using EBImage package for texture feature estimation, ROSE package for dataset balancing, and CARET package for training and testing the model.
Results: Of 77 LCH patients (mean age = 7.89 years with SD = 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 42 Tuberculosis patients (mean age = 32.6 years with SD= 12.5 years; range: 10 years to 64 years) 29 patients were male and 13 patients were female.
The best performing model was found to have 100% accuracy, 100% sensitivity and 100% specificity on training dataset. On test dataset the model had accuracy of 90%, sensitivity of 80%, and specificity of 100% in differentiating lytic bone lesions of LCH and Tuberculosis.
Conclusions: The machine learning model using Haralick texture feature of the lytic lesion and random forest algorithm as a classifier for differentiating lytic bone lesion of LCH and Tuberculosis was found to have 90% accuracy, 80% sensitivity and 100% specificity.