Feature extraction-based liver tumor classification using Machine Learning and Deep Learning methods of computed tomography images

Abstract

The liver is an important and multifunctional human organ. Early and accurate diagnosis of a liver tumor can save lives. Computed Tomography (CT) images provide comprehensive information for liver tumor diagnosis using feature extraction techniques. These extracted features help classify liver tumors using Machine Learning (ML) and Deep Learning (DL) methods. For this research, twelve 1hundred CT images were acquired from the Radiology Department of Nishter Medical University & Hospital. The noise was removed using Gabor Filter after converting CT images into grayscale. Image quality was enhanced by adopting Histogram Equalization (HE), and finally, the Image’s edges and boundaries were improved using a smoothening and sharpening algorithm. Preprocessed images were forwarded to extract six features: Histogram, Run-length, Co-occurrence, Autogressive, Gradient, and Wavelet Transform. A major focus of this research is to evaluate that ML methods produced good accuracy using already extracted features while DL Algorithms could not produce better results. Firstly, ML methods such as Decision Tree (DT), Random Forest (RF), Boost, and Support Vector Machine (SVM) are deployed using an already extracted feature list containing all six features. It was observed that DT, RF, Boost, and SVM produced 96.5%, 99.6%, 99.7%, and 98.0% classification accuracy. After that, DL Algorithms such as Neural Networks (NN), Long-Short Term Memory (LSTM), Bi-Directional Long Short Term Memory (Bi-LSTM), and Convolutional Neural Networks (CNN) were deployed. The results showed that NN, LSTM, Bi-LSTM, and CNN produced 50.0%, 53.0%, 54.0%, and 54.0% accuracy respectively. To validate the major focus of this research, finally, Pre-trained DL Algorithms such as Residual Network 50 (Resnet50), Visual Geometry Group 16 (VGG16) and LSTM + CNN were deployed. Results showed that Resnet50, VGG16, and LSTM + CNN attained 78.0%, 88.0%, and 97.0% accuracy respectively. Hence, ML methods performed better using already extracted features, while DL Algorithms could not produce promising results on these extracted features.