Deep learning based feature extraction for liver tumour classification using computed tomography images

Feature Extraction (FE) plays a vital role in medical image classification. FE methods convert raw pixel data into meaningful information such as shape, color, edges, textures, and object parts. Machine Learning (ML) algorithms use these features to resolve classification problems. Traditional FE methods are time consuming and complex, which affects the performance of ML algorithms. To address these challenges, there is a need to develop a framework that integrates Deep Learning (DL) feature extraction with classical ML algorithms. The focus of current research is to extract discriminative features from preprocessed Computed Tomography (CT) images using two pre-trained DL algorithms, such as AlexNet and GoogleNet, by removing their final classification layer. The extracted deep features were adopted by ML algorithms, namely Random Tree (RT), Random Forest (RF), Decision Tree (DT), Bagging, Hoeffding Tree (HT), and Naïve Bayes (NB), to perform binary classification of Tumour and non-Tumour. The dataset contains fourteen hundred CT images, equally distributed across two classes, namely Tumour and Non-Tumour. All the images were preprocessed for noise removal, histogram equalization, and grey scale conversion. Experimental results showed that ML classifiers produced better results using deep features. In particular, RF produced the highest classification accuracy of 99.93% using AlexNet features and 99.86% using GoogleNet features. The results indicate that deep feature extraction, when combined with ML classifiers, can produce an accurate and computationally efficient system without end-to-end deep network training requirements.