Automated classification of benign and malignant cells from lung cytological images using deep convolutional neural network

Background: Lung cancer is a leading cause of death worldwide, and its early detection is usually performed with low-dose computed tomography. For lesions suspected of abnormality by CT examination, a cytological diagnosis of lung cells collected by biopsy is first performed. However, atypical cells are the major challenge in malignant lung cell classification for cytotechnologists and cytopathologists. In this study, we aimed to automatize the classification of malignant lung cells from microscopic images using a deep convolutional neural network (DCNN). Method: Cytological specimens were prepared with a liquid-based cytology system and stained using the Papanicolaou technique. Images were acquired with a digital still camera attached to a microscope with a 40× objective lens. The original microscopic images were first cropped to obtain image patches with resolution of 224 × 224 pixels. We obtained 306 benign and 315 malignant image patches. To avoid overfitting, 60,000 patch images were generated using data augmentation by applying rotation, flipping, filtering, and color adjustment. DCNN classification was conducted based on a fine-tuned VGG-16 model. We performed patch-based segmentation of malignant regions in the images and evaluated classification performance using threefold cross-validation. Results: The classification sensitivity and specificity were 89.3 and 83.3%, respectively, reaching a performance comparable to that of a cytopathologist. Using the gradient-weighted class activation mapping, we visualized the DCNN identification performance while the network searched for typical benign and malignant cells in images for classification. Conclusions: The proposed method can be useful for accurate and automatic classification of lung cells from pulmonary cytological images.