Deep learning analysis for automatic lung nodule detection.

27 Background: The prevalence of lung cancer has been increased markedly in worldwide range with growing clinical significance, the quantitative and qualitative analysis on lung nodules has proven to be important for the early-detection of lung cancer as well as its treatment in clinical practice. However, lung lesion screening performed by radiologists can be very time-consuming and its accuracy varies depending on doctor’s individual experiences. In this study, we aim to build up a robust CAD system that automatically detects the lesion locations and quantitatively characterizes the detected lesions on CT images. Methods: Specifically, we employed the deep learning analysis for lesion detection in patients and performed image processing techniques to generate quantitative morphology features for assisting lesion diagnosis . The data collected includes 3956 lung CT series (slice thickness≤3mm) with multiple lung nodules from 15 Class-A hospitals in China , 1155 lung CT scan from Luna16 dataset as well as CT scans from Kaggle dataset (Data Science Bowl 2017). Lung nodule annotation was then performed by two experienced radiologists and further assessed by four senior associate chief physicians. The obtained CT images were randomly selected and split to construct training, validation and test dataset. After preprocessing, a pre-trained ResNet18 framework is transferred to develop a robust detection system to detect the possible lung lesion locations with corresponding probabilities. Results: The resulting detection system yields FROC of 0.4663, recall of 82.46%, precision of 36.06% for 5~30mm nodules. Each detected lesion was labeled by its bounding box and was then analyzed through image processing algorithm to generate diagnostic assisting features, including longest diameter, shortest diameter, volume, largest cross section area as well as its density type (calcify, solid, partial solid, and ground-glass opacity). Conclusions: The proposed CAD system offers a fast and convenient approach for assisting the diagnosis of lung nodule pathologies, and it is beneficial to relate our research to the current framework of lung cancer diagnosis.