pyMEAL A Multi-Encoder Augmentation-Aware-Learning Toolbox for Robust Medical Image Translation

Jul 16, 2026

Upload by Logic

Medical imaging is essential for clinical diagnosis, yet robust adoption of AI methods is hindered by patient variability, image artifacts, and limited generalization across acquisition settings. 3D tasks are further weakened by data scarcity and differences in scanners, protocols, and motion. Conventional augmentation uses a single transformation pipeline, ignoring augmentation-specific characteristics. MEAL uses multiple encoder pathways with three fusion strategies to integrate augmentation-aware features before decoding, with MEAL-BD using dynamic weighting for improved robustness.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages46

Reading Time1 h 56 min

Introduction

Method (MEAL)

Multi-encoder augmentation-aware design

Fusion strategies

Experiments

CT-to-T1 MRI translation setup

Results and metrics (PSNR/SSIM)

Discussion

arXiv :2505 .24421v2 [ ee ss .IV] 13 Jul 2026 pyMEAL: A Multi-Encoder Augmentation-Aware-Learning Toolbox for Robust Medical Image Translation Abdul-mojeed Olabisi Ilyas ∗1, Adeleke Maradesa †1, Jamal Banzi 1,5 , Jianpan Huang6,7,8 , Henry K.F. Mak6,7,8 , and Kannie W.Y. Chan‡1,2,3,4 1 Hong Kong Centre for Cerebro-Cardiovascular Health Engineering (COCHE), Hong Kong, China 2 Department of Biomedical Engineering, City University of Hong Kong, Hong Kong, China 3 Russell H. Morgan Department of Radiology and Radiological Science, The Johns Hopkins University School of Medicine, Baltimore, MD, USA 4 City University of Hong Kong Shenzhen Research Institute, Shenzhen, China 5 Department of Informatics, Sokoine University of Agriculture, Chuo Kikuu Morogoro, Tanzania 6 Department of Diagnostic Radiology, The University of Hong Kong, Hong Kong, China 7 State Key Laboratory of Brain and Cognitive Sciences, The University of Hong Kong, Hong Kong 8 Alzheimer’s Disease Research Network, The University of Hong Kong, Hong Kong Abstract Medical imaging is critical for clinical diagnosis, yet the adoption of advanced AIdriven imaging methods remains challenged by patient variability, image artifacts, and limited robustness across acquisition conditions. Although deep learning has transformed medical image analysis, 3D imaging tasks continue to suffer from data scarcity and variability arising from scanner differences, acquisition protocols, and patient motion. Conventional data augmentation typically relies on a single transformation pipeline, overlooking augmentation-specific characteristics and limiting effective representation learning. To address these challenges, we propose a Multi-Encoder Augmentation-Aware Learning (MEAL) framework, which processes multiple augmentation variants through dedicated encoder pathways. Three fusion strategies, encoder concatenation (CC), fusion layer (FL), and an adaptive controller block (BD), are introduced to integrate augmentationspecific features prior to decoding. Among these, MEAL-BD preserves augmentationaware representations through dynamic feature weighting, enabling improved robustness to clinically relevant variability. We evaluate MEAL in a CT-to-T1-weighted MRI translation case study, where such translation is particularly relevant in settings in which MRI is unavailable, contraindicated, or delayed, including emergency imaging and resource-limited clinical environments. Across unseen and predefined test data, MEAL-BD consistently outperformed competing approaches under both geometric perturbations and non-augmented conditions, achieving higher peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) . By prioritizing structural fidelity over perceptual realism, MEAL is designed to support clinical interpretation and downstream analysis rather than directly replacing diagnostic MRI, advancing robust and clinically meaningful medical image translation. Keywords: Computed Tomography, Magnetic Resonance Imaging, Augmentation-Aware Learning, Computer Vision, Artificial Intelligence ∗ Corresponding author: [amoilyas@hkcoche.org](amoilyas@hkcoche.org) †Corresponding author: [amaradesa@hkcoche.org](amaradesa@hkcoche.org) ‡Corresponding author: [kannie@hkcoche.org](kannie@hkcoche.org) 1 Introduction Medical imaging has revolutionized modern diagnostics, enabling precise disease detection [1] and treatment planning [2, 3] . Deep learning (DL) has played a transformative role in this domain, excelling in tasks such as lesion detection [4], segmentation [5, 6], and anomaly classification [7] . These advances are rooted in DL’s ability to learn hierarchical feature representations that align well with complex patterns in biomedical data [8, 9] . While DL has further enhanced these capabilities through automated image analysis, its application to 3D medical imaging remains affected by critical challenges such as limited availabil

FAQs

What problem does pyMEAL (MEAL) target in medical image translation?

MEAL targets poor robustness caused by patient variability, image artifacts, and differences in acquisition conditions, especially in 3D settings with limited and heterogeneous data.

How does MEAL incorporate augmentation information?

MEAL processes multiple augmentation variants through dedicated encoder pathways and integrates augmentation-specific features via fusion strategies before decoding.

What does MEAL-BD contribute to robustness and how is it evaluated?

MEAL-BD uses dynamic feature weighting to preserve augmentation-aware representations; evaluation in a CT-to-T1 MRI translation study shows improved performance under geometric perturbations and standard (non-augmented) conditions using PSNR and SSIM.

Jul 16, 2026

Upload by Logic