Transfer Learning by Deep Tuning of Pre-trained Networks for Pulmonary Nodule Detection
Dhaarna Sethi, Kriti Arora, Seba Susan
Abstract
Lung Cancer is one of the most common forms of cancer found worldwide and the accurate detection of lung nodules from computed tomography (CT) scans is a crucial preliminary step in the diagnosis procedure. In this paper, we address the problem of pulmonary nodule detection with an aim to design a robust model using deep convolutional neural networks (CNN) with transfer learning and fine-tuning of network weights. We use the benchmark LIDC/IDRI dataset which is derived from a set of CT scans that are cropped using annotations provided by a radiologist. Our study focuses on observing the effects of knowledge transfer from a non-medical domain to a medical domain using the pre-trained network architectures of VGG-19, ResNet-50, MobileNet and Inception-V3. We also analyze the effects of fine-tuning on the performance of the network. Shallow tuning refers to fine-tuning only the last few layers of the deep network while deep tuning refers to fine-tuning all the layers of the deep convolutional network. We compare the performance of the state-of-the-art convolutional architectures pre-trained on the ImageNet database, for both shallow and deep tuning.