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:

While a number of artificial intelligence (AI) algorithms exist for diagnostic tasks like nodule detection and classification, their clinical adoption is hindered by significant variability in computed tomography (CT) acquisitions, which impacts AI performance. Variations in scanner hardware, reconstruction methods, dose levels, and patient demographics cause domain shift, underscoring the need for harmonization and adaptation strategies to ensure consistent performance across diverse settings. To address this challenge, we propose a method that leverages knowledge of the AI model’s training distribution to guide the training of a normalizing flow-based harmonization method. The goal is to harmonize input images so that their distribution closely aligns with that of the downstream model's training domain. Our proposed approach, Cycle-CTFlow, leads to improvements in nodule detection performance: a 5.2% increase in sensitivity and a 2.6% increase in competition performance metric (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 diagnostic dose dataset.

Submission Number: 89