A feature extraction based support vector machine model for rectal cancer T-stage prediction using MRI images

Accurate clinical cancer T-stage diagnosis is crucial for effective treatment. However, it is difficult, time-consuming, and laborious for physicians to recognize T-stage manually using rectal Magnetic Resonance Imaging (MRI) images. Machine learning methods have played important roles in medical image processing. With the goal of automatic rectal cancer T-stage prediction, we train the proposed Feature Extraction based Support Vector Machine (FE-SVM) model with the newly acquired dataset consisting of 147 patients’ MRI images with primary rectal cancer. Our method adapts SVM as the training framework as SVM is effective enough for dealing with small datasets as opposed to state-of-the-art deep learning models. FE-SVM firstly extracts image similarity as an initial feature because the feature of image similarity can better reflect the differences among various types of MRI images, and the final 10-dimensional features are obtained by a 5-layers Autoencoder. To evaluate the performance of FE-SVM, we compared it with six benchmark models: CNN, Alexnet, Resnet18, Resnet50, Capsule Network, and Random Forest. FE-SVM outperforms these state-of-the-art models with significant evaluation scores.