Textural Classification of Multiple Sclerosis Lesions in Multimodal MRI Volumes

Background and objectives: Multiple Sclerosis is a common relapsing demyelinating disease causing the significant degradation of cognitive and motor skills and contributes towards a reduced life expectancy of 5 to 10 years. The identification of Multiple Sclerosis Lesions at early stages of a patient’s life can play a significant role in the diagnosis, treatment and prognosis for that individual. In recent years the process of disease detection has been aided through the implementation of radiomic pipelines for texture extraction and classification utilising Computer Vision and Machine Learning techniques. Eight Multiple Sclerosis Patient datasets have been supplied, each containing one standard clinical T2 MRI sequence and four diffusion-weighted sequences (T2, FA, ADC, AD, RD). This work proposes a Multimodal Multiple Sclerosis Lesion segmentation methodology utilising supervised texture analysis, feature selection and classification. Three Machine Learning models were applied to Multimodal MRI data and tested using unseen patient datasets to evaluate the classification performance of various extracted features, feature selection algorithm sand classifiers to MRI volumes uncommonly applied to MS Lesion detection. Method: First Order Statistics, Haralick Texture Features, Gray-Level Run-Lengths, Histogram of Oriented Gradients and Local Binary Patterns were extracted from MRI volumes which were minimally pre-processed using a skull stripping and background removal algorithm. mRMR and LASSO feature selection algorithms were applied to identify a subset of rankings for use in Machine Learning using Support Vector Machine, Random Forests and Extreme Learning Machine classification. Results: ELM achieved a top slice classification accuracy of 85% while SVM achieved 79% and RF 78%. It was found that combining information from all MRI sequences increased the classification performance when analysing unseen T2 scans in almost all cases. LASSO and mRMR feature selection methods failed to increase accuracy, and the highest-scoring group of features were Haralick Texture Features, derived from Grey-Level Co-occurrence matrices