Transfer Learning Vs. Fine-Tuning in Bilinear CNN for Lung Nodules Classification on CT Scans

Abstract

Lung cancer is one of the leading causes of death worldwide. Its early detection in its nodular form is extremely effective in improving patient survival rate. Deep learning (DL) and especially Convolutional Neural Network (CNN) have an important development over the past decade and were largely explored in medical imaging analysis. In this paper, a trending DL model composed of two CNN streams, named Bilinear CNN (B-CNN), was proposed for lung nodules classification on CT scans. In the developed B-CNN model, the pre-trained VGG16 architecture was trained as a feature extractor. It is the most important part of the proposed model in which its effectiveness depends stringently on its performances. Aiming to improve these performances, we address this question: what process leads with the performance improvement of the feature extractors? Transfer learning or Fine-tuning? To answer this question, two B-CNN models were implemented, in which the first one was based on transfer learning process and the second was based on fine-tuning, using VGG16 networks. A set of experiments was conducted and the results have shown the outperformance of the fine-tuned B-CNN model compared to the transfer learning-based model. Moreover, the proposed B-CNN model was demonstrating its efficiency and viability for the classification of lung nodules in terms of accuracy and AUC compared to existing works.