Cycle-CTFlow: A CycleGAN–Normalizing Flow Harmonization Framework for Improved Nodule Detection

Anil Yadav, Nathan Tran, Spencer Welland, John M. Hoffman, Grace Hyun J. Kim, Ashley E. Prosper, Denise R. Aberle, Michael McNitt-Gray, William Hsu

Keywords: Computed Tomography, Harmonization, Normalizing Flow

Abstract: Although many deep learning models exist for nodule detection, characterization, and other related tasks, their widespread clinical adoption is hindered by substantial variability in computed tomography (CT) acquisitions. Differences in scanner hardware, reconstruction methods, dose levels, and patient demographics cause domain shifts, often leading models trained on one dataset to underperform on another. This highlights the need for harmonization and adaptation strategies to ensure consistent performance across diverse clinical settings. To address this challenge, we propose a training scheme in which knowledge of the downstream model’s training distribution is incorporated into the training process of the harmonization model. The goal is to guide the harmonization model to transform input images so that their distribution closely aligns with that of the downstream model. We demonstrate that the proposed approach, Cycle-CTFlow, leads to improvements in nodule detection performance: a 5.2% increase in sensitivity and a 2.6% increase in CPM compared to no harmonization on the MiniDeepLesion dataset, and a 1.9% increase in sensitivity and an 8.5% increase in CPM on the UCLA in-house dataset.

Submission Number: 89