Classification of oesophagic early-stage cancers: deep learning versus traditional learning approaches

Abstract

Esophageal cancer is a disease with a high prevalence which can be evaluated by a variety of imaging modalities. Computer vision techniques could provide a valuable help in the analysis of these images, for it would allow an enhancement in diagnostic and staging accuracies, a decrease in medical workflow time and preventing patients’ loss of quality of life.Traditional learning techniques are frequently used in the biomedical imaging field, and deep learning algorithms are starting to see their rise in usage in this field as well. In this paper, both traditional and deep learning algorithms are applied on a dataset provided by Instituto Portugues de Oncologia (IPO) consisting of CT and three PET scans acquired at different treatment phases of 14 patients with oesophageal cancer.The main goal is to distinguish patients that need surgery from the ones that do not. The traditional learning method consisted of manually extracting the features and applying feature selection algorithms for further classification. Feature level and decision level fusion were also conducted. The deep learning method consisted of using convolutional neural networks to extract and classify the image features. Moreover, traditional and deep learning techniques were used simultaneously, where the features were extracted and selected by a pretrained network and classified using the traditional learning classifiers.Traditional Learning methods achieved 92.86% accuracy, while for feature extraction with deep learning followed by classification with a traditional classifier was able to reach 100% accuracy. The difference has, however, proven not to be statistically significant. In this way, for this particular problem and conditions, it can be said that traditional techniques are capable of achieving results as good as with deep learning.