Faster R-CNN based on Optimized SqueezeNet for Liver Lesion Detection from DeepLesion Dataset

Abstract

Computed tomography (CT) images have been widely used by radiologists and doctors to diagnose patients with liver cancer. However, liver lesion detection from CT images can be difficult due to the similar characteristics of the lesion region and the surrounding region. The recent developments in deep learning methods including deep learning object detection models, such as Faster Region with Convolutional Neural Network (Faster R-CNN) has shed light on whether it can contribute to liver lesion detection from CT images. Until recently, the pre-trained deep neural networks, ResNet50 and ResNet101, are applied as the feature extractors for Faster R-CNN to detect liver lesions in CT images. Nevertheless, Faster R-CNN with ResNet50 and ResNet101 have produced a very low average precision (AP) and large network size due to their massive number of parameters. In this research, a lightweight architecture has been implemented based on SqueezeNet and Faster R-CNN (Squeeze-Faster R-CNN) to detect liver lesions from CT images. The architecture of SqueezeNet has been optimized by reducing the depth of the network from 18 to 15, the number of filters has been decreased from 96 to 64, and the size of the filters was reduced from 7 × 7 to 3 × 3 in the first convolutional layer. This has improved the AP from 42.7 to 63.8 and has further reduced the number of parameters from around 1.24 million to 0.53 million. This optimized framework can be beneficial for lesion detection on embedded devices with limited memory.