Automatic prognosis of lung cancer using heterogeneous deep learning models for nodule detection and eliciting its morphological features

摘要

Among cancers, lung cancer has the highest morbidity, and mortality rate. The survival probability of lung cancer patients depends largely on an early diagnosis. For predicting lung cancer from low-dose Computed Tomography (LDCT) scans, computer-aided diagnosis (CAD) system needs to detect all pulmonary nodules, and combine their morphological features to assess the risk of cancer. An automatic lung cancer prognosis system is proposed. The existing CAD system is only for nodule detection. Actually, presence of a nodule does not mean cancer. Depending on its morphological features, the risk that it eventually would develop into cancer, is different. The motivation of the work is to propose a complete lung cancer prognosis system. It consists of 2 cascaded modules: nodule detection module and cancer risk evaluation module. In nodule detection module, two object detection algorithms are ensembled to minimize missing detection, i.e., maximize recall performance. They are based on 3D convolutional neural network (3D-CNN), and our recently proposed model of recurrent neural network (RNN). As they extract features in completely different ways, we call them heterogeneous deep learning models. By ensembing them, we could achieve much better recall performance compared to individual detectors. In cancer risk evaluation module, 3D-CNN based models are trained to evaluate the grade of malady of morphological features of pulmonary nodules. It will also provide medically interpretable intermediate information. Finally, a regression model is trained to match the ground truth labels describing morbidity grade of the CT-Scan. In this work, 13 features from the highest risk nodule is used to evaluate the risk of lung cancer. We also identify the subset of structural and morphological features which are strongly related to grading decision, labelled by oncologist. The final system could obtain a low logloss of 0.408.