Cross-Institutional Generalization in Medical Deep Learning

494 concepts465 papers29 proposals

Imagine training a state-of-the-art deep learning model that achieves near-perfect accuracy at detecting tumors on MRI scans from your local hospital. Confident in its performance, you deploy this model at a clinic across town. Instantly, its accuracy plummets. The model hasn't forgotten how to read scans; rather, it has fallen into the trap of single-center data bias.

In medical machine learning, models frequently overfit to the hidden, non-clinical signatures of the specific institution where they were trained. Achieving true clinical utility requires closing the multi-center validation gap.

The Single-Center Trap

When a machine learning model is trained on data from a single medical facility, it doesn't just learn the biological markers of a disease. It also inadvertently learns the quirks of that specific facility's environment. For example, Yanhui Feng et al. developed an XGBoost pipeline using MRI radiomics that achieved an outstanding AUC of 0.905 for discriminating testicular lesions (Comparison and analysis of multiple machine learning models for discriminating benign and malignant testicular lesions based on magnetic resonance imaging radiomics). However, the authors explicitly noted that such single-facility models often lack robustness when confronted with different patient demographics or imaging equipment.

This phenomenon occurs because neural networks are opportunistic shortcut learners. If a specific scanner model or clinical protocol correlates slightly with a patient outcome, the network will exploit that feature to minimize training loss. When the model is moved to a new institution, these shortcuts vanish, leading to catastrophic performance degradation.

The Anatomy of Center Bias

What exactly constitutes this "center bias"? The variations across medical institutions can be categorized into three main dimensions:

1. Hardware and Acquisition Protocols

Different manufacturers (e.g., Siemens, GE, Philips) use distinct reconstruction algorithms, sensor technologies, and calibration settings. As BIN CHEN et al. point out, variations in imaging hardware and scanning protocols across multiple centers directly degrade the performance of predictive algorithms (Computed Tomography Radiomic Nomogram for Preoperative Prediction of Extrational Extension in Papillary Thyroid Carcinoma). Even within the same imaging modality, changes in contrast agent timing or slice thickness alter the image texture. Guotai Wang et al. observed that training on single-institution data severely limits a model's ability to handle variations in image quality and contrast produced by different MRI scanners (Automatic Segmentation of Vestibular Schwannoma from T2-Weighted MRI by Deep Spatial Attention with Hardness-Weighted Loss).

2. Sample Preparation and Staining

In digital pathology, the bias is often chemical. Young Ki Lee et al. emphasize that variations in laboratory hardware, sample preparation techniques, and local clinical protocols prevent single-center models from generalizing (Machine-learning-based diagnosis of thyroid fine-needle aspiration biopsy synergistically by Papanicolaou staining and refractive index distribution). A tissue slide stained in one lab may have a slightly warmer hue or higher contrast than one prepared elsewhere. Wenhui Wang et al. confirmed that variations in staining equipment and chemical reagents across different laboratories create severe domain shifts, causing models to struggle when deployed in new clinical environments (Quantitative immunohistochemistry analysis of breast Ki67 based on artificial intelligence).

3. Demographic and Selection Bias

The patient population of a tertiary referral hospital in a major metropolitan area looks vastly different from that of a rural community clinic. Differences in age, ethnicity, and disease severity shift the underlying data distribution, rendering single-center models fragile.

Exposing the Illusion of High Accuracy

Without testing a model on external data, its reported accuracy can be dangerously misleading. In a systematic review of machine learning in neuroimaging, Enrico Pellegrini et al. found that a heavy reliance on single, centralized datasets creates a pervasive risk of inflated accuracy estimates (Machine learning of neuroimaging to diagnose cognitive impairment and dementia: a systematic review and comparative analysis). The models appear highly accurate because they are evaluated within the same homogeneous environment they were trained in.

To combat this, researchers are shifting toward multi-center validation as a mandatory benchmark. By evaluating a model on completely unseen data from independent institutions, developers can measure true generalizability. Julius C. Holzschuh et al. explicitly quantified this effect by analyzing PET scans across four distinct institutions, demonstrating that single-center datasets cause models to overfit to specific institutional workflows and hardware configurations (The impact of multicentric datasets for the automated tumor delineation in primary prostate cancer using convolutional neural networks on 18F-PSMA-1007 PET).

Why Multi-Center Training Works

The most effective solution to single-center bias is integrating multi-institutional data directly into the training phase. Understanding why this works requires looking at how neural networks construct internal representations.

When a network is exposed to data from only one scanner, the scanner's specific artifacts are static constants. The network incorporates these constants into its decision boundaries. However, when you introduce data from multiple centers, those scanner-specific artifacts effectively become noise.

Because the artifact features are no longer predictive of the target label across the entire training set, the network is forced to discard them. To minimize loss across diverse data sources, the model must locate features that remain invariant across all institutions—namely, the actual anatomy and pathology of the patient.

This push for diverse data curation spans across all medical modalities:

By breaking the reliance on single-center data, the medical AI field moves away from fragile, site-specific algorithms and closer to robust, universally deployable diagnostic tools.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 379 papers
  2. Version 1 · 2026-10-04 · 327 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_single_center_data_bias_8331582f,
  title = {single-center data bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-center-data-bias/rdr_8331582f91a3482990840454ea89edcd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_single_center_data_bias_8331582f_v2,
  title = {single-center data bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-center-data-bias/rdr_8331582f91a3482990840454ea89edcd/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake