Biologically Inspired Medical Multi-Modal Dataset Distillation via Contrast-Aware Alignment and Memory Compression.

Overview

1. College of Computer Science and Technology, Jilin University, Changchun 130012, China; (T.D.); (Z.W.); (Y.W.); (M.M.)
2. School of Games, Jilin Animation Institute, Changchun 130013, China

Abstract

Multi-modal Magnetic Resonance Imaging (MRI) provides complementary information for clinical diagnosis, yet its large-scale storage, privacy sensitivity, and annotation cost pose significant challenges. Inspired by biological vision systems, which integrate multi-sensory inputs and compress experiences into compact memory representations, we propose a bio-inspired framework termed Contrast-Guided Multi-modal Dataset Distillation (CGMDD). In biological perception, different sensory channels observe the same environment from complementary perspectives, while hierarchical neural processing ensures perceptual consistency across modalities. Meanwhile, memory systems such as the associated medial temporal lobe structures consolidate redundant experiences into efficient representations for long-term storage. Motivated by these principles, CGMDD treats multi-modal MRI as multi-view perceptual signals and introduces a hierarchical cross-modal contrastive learning mechanism that enforces perceptual alignment across modalities, analogous to multi-level processing in the visual cortex. Furthermore, we design a dynamic dataset distillation strategy that mimics memory consolidation by compressing large-scale data into compact, informative synthetic representations through gradient-based optimization. The proposed framework jointly optimizes perceptual alignment and memory compression in an end-to-end manner, achieving a biologically plausible integration of perception and learning. Experimental results on two MRI datasets demonstrate that CGMDD can compress the original dataset to 5% of its size while maintaining competitive performance, even with only 30% of the labels. These findings highlight the effectiveness of bio-inspired mechanisms in building efficient, robust, and privacy-preserving computer vision systems.

Reproduced under the paper's license (CC BY), from the paper cited above.

Code

The paper links to its data, not to its authors' code: see the Data section.

The paper's code and data availability statement is in the Data section.

Tracing map

A tracing map links a paper to the code its authors published: this paper has none, so it has no map.

Data

Datasets cited

• github.com/wjx0818, at github.com; found in “Data Availability Statement”

2.3. Medical Data Sharing and Privacy Protection

To address the challenges of data scarcity and privacy sensitivity in medical imaging, researchers have proposed a variety of solutions, including data augmentation (e.g., random transformations and generative models such as GANs or diffusion models) [48,49,50,51], transfer learning (fine-tuning pre-trained models) [52], lightweight network design (e.g., 3D variants of MobileNet and ShuffleNet) [53], and federated learning (privacy-preserving distributed training). While effective in specific contexts, these methods do not fundamentally reduce reliance on the original datasets. Data augmentation emphasizes data expansion rather than compressing [54,55]; transfer learning still depends on real samples from the target task for fine-tuning; lightweight models reduce computational overhead often at the expense of performance; and federated learning, despite its data privacy benefits, still requires complete data support locally. In contrast, dataset distillation offers a fundamentally different solution by directly compressing the data itself. It significantly reduces the cost of computation, storage, and transmission while maintaining competitive performance. Unlike the method of removing identifiers to protect patient privacy [56], or the privacy-preserving medical data sharing schemes in cloud environments proposed by Yang et al. [57] and Fabian et al. [58], which only achieve superficial anonymization, dataset distillation eliminates potential privacy leakage risks at the data generation level. It enables genuine, complete anonymization of medical image data itself, thereby fundamentally addressing privacy protection—the primary barrier to secure medical data sharing [59].

Our work, in response to the aforementioned limitations and inspired by the hierarchical perception and multi-view integration mechanisms of the biological vision system, proposes a unified framework that integrates multi-view visual consistency and efficient representation learning into dataset distillation. Specifically, biological vision systems process multiple visual observations of the same scene through hierarchical structures in the visual cortex, where complementary information is progressively aligned and integrated into coherent representations. Motivated by this mechanism, our work takes multi-modal medical image feature learning and fusion as the foundation and designs a hierarchical contrastive learning strategy to explicitly model cross-modal consistency.

Furthermore, a key characteristic of biological vision lies in its ability to compress redundant visual inputs into compact and efficient representations through efficient coding mechanisms, enabling long-term retention with minimal redundancy. Inspired by this principle, we reinterpret dataset distillation as a form of visual information compression and incorporate the objectives of cross-modal contrastive learning into the distillation optimization process in a novel manner. This design allows the distilled data to explicitly preserve complementary information across modalities while reducing redundancy, thereby guiding the generation of compact yet informative synthetic samples. While CGMDD is inspired by biological vision and memory systems, we emphasize that our goal is not to faithfully replicate neurobiological processes, but to derive functionally analogous computational principles. Specifically, certain components in our framework are designed as functional abstractions of biological mechanisms. For example, the hierarchical cross-modal contrastive learning scheme reflects the progressive alignment of multi-level representations in the visual cortex, while the dynamic dataset distillation process is inspired by memory consolidation mechanisms that compress redundant experiences into compact representations. These correspondences are not intended to be exact biological implementations. Our framework does not explicitly simulate neural circuitry or temporal dynamics in biological systems. Instead, it adopts biologically motivated design principles to guide the development of efficient and robust learning algorithms. As a result, the proposed approach achieves a balanced trade-off among compression efficiency, diagnostic performance, and privacy protection.

Reproduced under the paper's license (CC BY), from the paper cited above.

Data Availability Statement

The MRNet and JLURM dataset presented in the study are openly available at the following URLs: https://stanfordmlgroup.github.io/competitions/mrnet/ (accessed on 9 January 2019) and https://github.com/wjx0818/ (accessed on 27 August 2025).

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 28 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 5 authors, 5 keywords, 1 funder, 34 references.

Cite

This paper

Du, T., Wang, Z., Wang, Y., Ma, M., & Li, W. (2026). Biologically Inspired Medical Multi-Modal Dataset Distillation via Contrast-Aware Alignment and Memory Compression. Biomimetics (Basel, Switzerland), 11(5), 314. https://doi.org/10.3390/biomimetics11050314

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.3390/jimaging12060255

A Multimodal Dense Parallel Global Attention Mechanism for Brain Tumor Image Segmentation.

Journal: Journal of imaging

In common: methods / tools, 1 reference

[2] doi:10.1186/s12911-026-03479-0 [code]

Synthetic cervical spine radiographs: expert validation and transfer learning for low-data settings.

Journal: BMC medical informatics and decision making

In common: methods / tools, 1 reference

[3] doi:10.3389/fradi.2026.1785108 [code]

BT-CAP: a subcomponent-aware and anatomically constrained data augmentation framework for multi-modal brain tumor MRI segmentation.

Journal: Frontiers in radiology

In common: methods / tools, 1 reference

[4] doi:10.3389/fonc.2026.1806663

A denoising autoencoder-enhanced ResNet 50 framework for explainable brain tumor classification using MRI.

Journal: Frontiers in oncology

In common: methods / tools, 1 reference

[5] doi:10.3389/frai.2026.1771088 [code]

Few-shot deployment of pretrained MRI transformers in brain imaging tasks.

Journal: Frontiers in artificial intelligence

In common: methods / tools, 1 reference

[6] doi:10.3389/frai.2026.1849571

NeuroPlast: a learnable activation function evaluated under knowledge distillation for medical image classification.

Journal: Frontiers in artificial intelligence

In common: methods / tools, 1 reference

[7] doi:10.3390/molecules31111884 [code]

Drift-Robust Lightweight Deep Learning on Open Gas Sensor Benchmarks: A Reproducible Architecture Study with CBRN Applicability Mapping.

Journal: Molecules (Basel, Switzerland)

In common: methods / tools, 1 reference

[8] doi:10.3390/diagnostics16111588 [code]

Bridging Annotation Gaps: Hierarchical Self-Support Learning for Brain Tumor Segmentation.

Journal: Diagnostics (Basel, Switzerland)

In common: methods / tools, 1 reference

[9] doi:10.3390/e28030310

Entropy-Based Dual-Teacher Distillation for Efficient Motor Imagery EEG Classification.

Journal: Entropy (Basel, Switzerland)

In common: methods / tools, 1 reference

[10] doi:10.1038/s41597-026-07248-6 [code]

A large-scale fMRI dataset for vision-language semantic association.

Journal: Scientific data

In common: methods / tools, 1 reference

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.