核心概念
Leveraging 3D Normalizing Flows for Unsupervised Detection of Pathological Pulmonary CT Scans.
統計資料
The dataset consists of 570 normal and 252 abnormal CT scans.
The model trained on 500,000 normal 3D CT patches.
The model architecture includes L = 4 blocks and K = 64 flows.
The network trained for 50,000 iterations on 2 NVIDIA A100 SXM4 GPUs.
引述
"Our 3D patch-based NF model demonstrates the superiority of 3D flow-based model over state-of-the-art 2D methods."