Automated lung cancer detection using histopathological images

Lung cancer is the leading cause of all cancer deaths. Assessment of histopathological images by a pathologist is the gold standard for lung cancer diagnosis. However, the number of qualified pathologists is too small to meet the substantial clinical demands. This study aimed to develop an automated lung cancer detection framework using while-slide histopathology images. The algorithm development consisted of the data splitting, data preprocessing, deep learning models development, training and inference processes. Two different U-Net variants (U-Net and U-Net++) with two different encoders (ResNet34 and DenseNet121) were selected as base models, and two loss functions including dice loss and weighted binary cross entropy loss were used during training. Unweighted average was used to combine results of multiple base models. On the test dataset, the ensemble model using 5X magnification and 512X512 patches obtained an accuracy, sensitivity, specificity and dice similarity coefficient of 0.934, 0.877, 0.948, 0.840, respectively. Except for the specificity of 10X magnification being slightly higher than that of 5X magnification, no matter what model type, encoder, loss function and performance metric were used, the performances of using the 5X magnification outperformed those of using the 10x and 20x magnifications. This algorithm achieved satisfactory results. Moreover, extensive experiments indicated that using 5X magnification 512X512 patches is a good choice in automated lung cancer detection. In the future, after improving the generalizability of this framework in real clinical settings, this framework can be used to assist histologists in their daily work.