Addressing Domain Shift in Medical AI: From Wavelet-Based Fourier Generalization to Efficient Pixel-Space Harmonization
ECE 595 seminar series
February 20, 2026
11:00 AM - 12:00 PM
Location
Lecture center C1
Address
802 S. Halsted St., Chicago , IL 60607
Calendar
Download iCal FileAddressing Domain Shift in Medical AI: From Wavelet-Based Fourier Generalization to Efficient Pixel-Space Harmonization
Presenter: Hongyi Pan, Northwestern University
Abstract: Clinical deployment of AI is often challenged by domain shifts arising from varied imaging protocols and hardware vendors. In this seminar, I present an evolution of methods to mitigate these shifts, starting with the application of wavelet-inspired thresholding to improve Fourier-based domain generalization (FDG). I will discuss the transition from soft-thresholding, which can introduce bias in large coefficients, to a more robust hard-thresholding approach integrated into a Federated Learning framework to preserve data privacy. Finally, I will introduce LUMINA, a multi-vendor mammography benchmark that demonstrates how the core objectives of frequency-domain alignment can be achieved more efficiently through foreground-only histogram matching. This progression highlights a shift toward faster, simpler, and more effective pixel-space harmonization techniques that enhance the reliability of medical AI in real-world clinical environments.
Speaker bio: Hongyi Pan is a postdoctoral research fellow in the department of radiology at Northwestern University. A proud alumnus of the University of Illinois Chicago, he received both his MS (2019) and PhD (2023) in electrical and computer engineering under the supervision of Ahmet Enis Cetin. His doctoral training focused on bridging classical signal processing with modern deep learning, where he developed novel neural network layers inspired by orthogonal transforms such as DFT, DCT, and Walsh-Hadamard Transform to enhance model robustness and spectral efficiency. During his PhD, he published as first author in top venues including ICASSP, EUSIPCO, ICPR, ICML, CVPR, ACM TESC, and IEEE TNNLS.
Since joining Northwestern under the supervision of Ulas Bagci, Pan has pivoted to medical AI, leading high-impact interdisciplinary research. He led the development of Cyst-X, an AI framework for pancreatic cyst risk prediction currently under review at Nature Communications, and he currently leads a major project on LUMINA for breast cancer analysis. His postdoctoral work has been featured in leading venues such as ICASSP, ISBI, and Medical Image Analysis. He also serves as an associate editor for the journal Signal, Image and Video Processing.
Faculty host: Pai-Yen Chen, pychen@uic.edu
Date posted
Feb 17, 2026
Date updated
Mar 6, 2026