Federated learning framework aims to stabilize colorectal cancer CT classification across hospitals

Researchers have introduced FedTDA-MAR, a personalized federated learning framework designed to classify colorectal cancer from CT images without moving patient data between institutions. The work, published in Cluster Computing, addresses a persistent obstacle in medical artificial intelligence: privacy constraints across multi-center datasets prevent centralized data sharing, which limits how well deep learning models generalize beyond a single hospital’s patient population.

The framework combines two mechanisms. First, a Transformer-based module models the structural and semantic relationships among client classifiers, producing differentiated aggregation weights for more precise parameter fusion. According to the authors, existing personalized federated learning methods often overlook the dynamic evolution of client importance during training, treating participating sites as if their relevance to the shared model were fixed. Second, a multi-component regularization strategy dynamically adjusts parameter constraints using both historical consistency and performance trends, aiming to counter parameter drift that can compromise model stability and personalization.

The challenge is that data from different hospitals are typically non-independent and identically distributed: scanner types, patient demographics and disease prevalence vary, so a single global model may perform well at some sites and poorly at others. Personalized federated learning partially addresses this heterogeneity, but the authors argue that current approaches remain vulnerable to the instability and drift problems their new aggregation and regularization scheme is designed to mitigate.

In experiments on colorectal cancer CT image classification, the authors report that FedTDA-MAR markedly improves model stability and generalization in heterogeneous environments. The published abstract describes the framework as integrating structure-aware dynamic aggregation with multi-level adaptive regularization, with the Transformer module producing aggregation weights that differ across clients rather than applying a uniform fusion rule.

The findings come from the authors’ reported experiments on CRC CT image classification, and the full details of the evaluation, including datasets, baselines and performance metrics, are available in the subscription article. As with any computational study, the reported improvements describe performance on the evaluated classification task and do not by themselves establish clinical deployment readiness. The work, authored by Zihao Zhao, Dinghui Wu, Yuxi Ge and Shudong Hu, appears in Volume 29 of Cluster Computing as article number 879.

Subject of Research: A personalized federated learning framework for colorectal cancer CT image classification

Article Title: FedTDA-MAR: federated learning with transformer-based dynamic aggregation and multi-adaptive regularization for colorectal cancer CT image classification

Article References: Zhao, Z., Wu, D., Ge, Y., & Hu, S. (2026). FedTDA-MAR: federated learning with transformer-based dynamic aggregation and multi-adaptive regularization for colorectal cancer CT image classification. Cluster Computing, 29(15), Article 879. https://doi.org/10.1007/s10586-026-06647-z

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06647-z

Keywords: federated learning, colorectal cancer, CT imaging, Transformer, deep learning, personalized federated learning, model aggregation, regularization, medical imaging, computer-aided diagnosis, privacy, Cluster Computing