Meet The Author
January 2026
Wilfried Mai DVM, MS, PhD, DACVR, DECVDI
Hi everyone! I’m Wilfried Mai, a professor of veterinary radiology at the University of Pennsylvania. Much of my clinical and research work focuses on cross-sectional imaging, particularly MRI. Aside from working as a clinical scientist and mentoring residents (42 of them during my tenure at Penn Vet thus far), I like to travel, cook, enjoy the fabulous restaurant scene in Philadelphia, and practice Brazilian Jiu-Jitsu (got my Black Belt two years ago).
A Veterinary DICOM-Based Deep Learning Denoising Algorithm Can Improve Subjective and Objective Brain MRI Image Quality
W. Mai, S. Hecht, M. Paek, S.P. Holmes, H. Dorez, M. Blanchard, J.N. Eddin
The Study Background
I am interested in how artificial intelligence can support veterinary radiologists, not replace them, by improving the clarity and utility of the images we interpret. This interest led our multidisciplinary team to explore whether a veterinary-specific deep learning denoising algorithm could meaningfully improve brain MRI image quality. MRI provides excellent soft-tissue contrast; however, veterinary MRI exams can be lengthy and are often limited by a low signal-to-noise ratio (SNR). Shortening scan time typically results in increased noise, decreased contrast, and reduced diagnostic confidence. Commercial deep-learning denoising tools exist for human MRI, but they are scanner-specific, often operate in k-space, and are trained entirely on human datasets.
HawkAI, a DICOM-based denoising system trained entirely on canine and feline MRI datasets from multiple hospitals and scanners, was the focus of this study. Because it is purely DICOM-based, it can work with any scanner and any protocol for real-world clinical patients.
What is the primary knowledge gap your study aims to address?
We lacked validated AI denoising methods specifically trained on the anatomic and contrast characteristics of dog and cat brain MRI images. Our study aimed to quantitatively and qualitatively evaluate the effect of this denoising algorithm on clinical brain MR images.
The Study Design
We evaluated 30 clinical brain MRI studies (dogs and cats) obtained on a GE Healthcare 1.5 T scanner using routine protocols.
Quantitative assessment: we compared native vs. AI-denoised images using:
Signal-to-noise ratio (SNR) in the gray matter, white matter, deep nuclei, and internal capsule.
Contrast-to-noise ratio (CNR) between key tissue pairs.
Sequences evaluated: T2W, T2-FLAIR, and Gradient Echo (GRE).
Qualitative assessment: three blinded ACVR board-certified radiologists used a standardized 0-3 scoring system and independently graded:
Coarseness (noise).
Perceived contrast.
Overall image quality.
The Algorithm (HawkAI):
Generative Adversarial Net-type model.
Trained on 34,841 paired canine and feline MRI images (clinical MRI scans of normal and diseased cats and dogs acquired from five different veterinary hospitals using both 1.5 T and 0.25 T magnets).
Works in image space, meaning it can be applied to images from any scanner.
What are the main study results?
1. Objective Image Quality Improved Significantly across all sequences. SNR increased (T2W: +88–94%; T2-FLAIR: +35–41%; GRE: +38–42%) and CNR increased (T2W: +100–105%; T2-FLAIR: +44–49%; GRE: +38–46%). These are substantial improvements — comparable to, and in some cases exceeding, proprietary human MRI denoising tools.
2. Radiologists Viewed the Images as Better. The blinded reviewers generally rated AI-processed images as having less coarseness, better contrast between brain structures, and higher overall image quality. This was most prominent in T2W and GRE sequences. T2-FLAIR improved as well, though subjective differences were sometimes subtler — likely due to its inherently grainy nature.
Were there any unexpected results or challenges during your research?
Even though the numerical differences between native and AI-processed images were less dramatic with T2-FLAIR than T2W or GRE, radiologists noticed subjective quality improvements. The largest SNR/CNR gains occurred in T2W, a notoriously signal-starved yet clinically important sequence for lesion detection.
Takeaways from this study
The HawkAI algorithm improves both subjective and objective brain MRI image quality for dogs and cats. It is the first veterinary-specific, DICOM-based, platform-agnostic deep learning denoising solution. These gains could: enhance detection of subtle lesions, allow shorter scan times (fewer averages), improve low-field MRI performance, and increase diagnostic confidence. Importantly, the algorithm does not require special hardware or vendor-locked software.
What future directions would you like to explore based on this study?
We see potential for side-by-side diagnostic accuracy studies, comparison with commercial manufacturer-specific AI reconstruction algorithms, application to additional pulse sequences, expanding to low-field MRI systems, and using denoising to enable faster MRI protocols in clinical practice.
