Detection of Lung Cancer from Pathological Images Using CNN Model

Over the past decade, Cancer detection using deep learning models has been a hot topic, especially in medical image classification. It is worth remarking that CNN models are more advanced at addressing diagnose diseases such as lung cancers because of the higher performance and ability of the CNN. However, few methods focus on lung pathological image classification and provide a detailed model configuration. In this paper, a hybrid classification model including inception_v3 network, hog and daisy feature extraction modules are constructed to classify lung cancer and normal tissue from lung pathological images. The work is carried out on Lung and Colon Cancer Histopathological Image dataset on Kaggle which includes three classes of lung images: normal, Adenocarcinoma and Squamous Cell Carcinoma with 5000 images in each class. In the classification process, 80% of the data is chosen by KFold as training set and fed into the hybrid model, in which features extracted by neural network are combined with that extracted by hog and daisy method. The experimental results show that an accuracy of 99.60% can be achieved in the proposed model, which is a satisfying result higher than other similar classification experiments. This outcome of the experiment seems to show that hybrid deep learning models can verify an accurate method of cancer diagnosing.