Harmonization for a black-box deep learning model

Deep learning has demonstrated remarkable success across various fields, leading to the deployment of numerous models in practical and commercial applications. However, most commercially available tools are a black-box model, where model parameters are inaccessible, preventing additional training or adaptation of this model to a new domain. MR images acquired from diverse environments (e.g., vendors, scanners, parameters) can contain domain gaps, which may hinder model performance on data from a different domain. To address performance degradation, studies have proposed harmonization methods. However, these methods often require access to target domain data or model parameters, limiting their applicability to a black-box model. To address this challenge, we propose BboxHarmony to train a harmonization network for a black-box model using a small dataset of source domain images and labels.

Authors: Minjun Kim, Hwihun Jeong, Hoigi Seo, Se Young Chun, and Jongho Lee