Transfer learning with GoogLeNet for detection of lung cancer

Abstract

The use of computer algorithms has gained momentum in filling/assisting roles of specialists especially in early diagnosis scenarios. This paper proposes the employment of deep neural networks (DNN) to detect images with malignant nodules of lung computed tomography (CT). The method includes subjecting input images to a simple and fast pre-processing which isolates regions of interest (ROI), that’s the lungs dominated area, ridding the images of other surrounding tissues and artefacts. Centered and size normalized images are then fed to a deep neural network for training and validation. In this work transfer learning is used to readjust GoogLeNet DNN to learn this medical data. This includes allowing final layers of the DNN to evolve while restricting deep layers. In this setting, a rough, unprocessed dataset, the IQ-OTH/NCCD lung cancer dataset was used to train/validate the proposed algorithm. Experimental results show that this algorithm scores 94.38% accuracy, which outperforms benchmark method previously used with this dataset.