ONCOLENS – Agentic Federated Learning for Cancer Detection Using Histopathology and MLOps
Histopathological image classification can support the analysis of tissue patches, but collaborative model development across institutions is constrained by data-governance requirements and the operational complexity of maintaining machine-learning systems. Federated learning enables participating clients to train a shared model while retaining their raw training images locally; however, it does not by itself provide update validation, experiment traceability, explainability, or model-lifecycle governance. This paper presents ONCOLENS, a research prototype that brings together ResNet-18-based binary histopathology classification, federated optimization, Grad-CAM visualization, a Django REST inference service, a React interface, and MLOps-oriented monitoring. The design considers FedAvg as the principal aggregation approach and FedProx and FedBN as comparison methods. The project uses the PatchCamelyon (PCam) benchmark and an IID client-partition configuration. The current project report records preliminary accuracy values of 94.13% for ResNet-18, 95.35% for fine-tuned ResNet-18, 95.09% for FedAvg, 95.05% for FedProx, 91.77% for FedBN, and 94.33% for DP-FedAvg. These figures are reproduced as project-reported values only; they have not been independently verified against final checkpoints and evaluation logs, and earlier project records contain conflicting values for some methods. Accordingly, they must not be treated as confirmed experimental findings until re-evaluated using a common held-out test set. The paper also distinguishes federated data locality from formal privacy guarantees and simulated clients from real clinical institutions. ONCOLENS is therefore presented as an engineering research prototype, not a clinically validated diagnostic system.
Authors
- S Hemalatha (ORCID: https://orcid.org/0009-0009-2016-7688)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24219720
- Primary Topic
- Privacy-Preserving Technologies in Data
- Type
- preprint