Automatic diagnosis of adrenal gland nodule via two-stage deep neural network

Abstract

Segmentation or Detection on CT images is a crucial for automatic diagnosis using deep learning on medical image. However, recent researchers only focus on detection of the lesion or nodule, not on the diagnosis level. Therefore, we propose an automatic diagnosis system that can detect adrenal gland organs and nodules on the CT images and to determine whether the patient has a nodule or not. Our model has a backbone similar to Mask R-CNN, a 3D feature fusion module to deal with consecutive CT images, and a two-stage framework to diagnose the nodularity. At the first stage, our model detect a region of adrenal gland or adrenal. The second stage model refine the detected region to segment the nodule. Then, we post-process the second stage output region to diagnosis the presence of absence of a nodule at the patient-level. We achieve the state-of-the-art segmentation performance on our Datasets and the diagnosis accuracy at the patient-level. We also conduct an ablation study and qualitative analysis of our results on our large datasets.