Hybrid one-shot and vertical federated learning framework for privacy-preserving brain tumor classification and segmentation using the BRISC2025 dataset

Introduction From MRI images, diagnosis of brain tumors may have limitations due to restriction over patient privacy and lack of annotated datasets so that models cannot be trained across multiple institutions. Methods HOS-VFL: A Hybrid One-Shot and Vertical Federated Learning framework with a two-round protocol that jointly enjoys the advantages of prototype-based one-shot aggregation and encrypted vertical feature space augmentation. Tested on BRISC2025 MRimages of glioma, meningioma, pituitary and non-tumor, with T1-weighted images, expert-annotated, at five simulated institutions, with simulated IID and Dirichlet setups. Results HOS-VFL obtained values of 98.73% on accuracy and 35–50 for conventional federated learning, with a 94.3% reduction compared to the best baseline valued. HOS-VFL performed 98.73% accuracy and 35–50 for the conventional federated learning with an improvement of 94.3% when compared with its best baseline accuracy. The mean Dice and IoU of Segmentation was 0.9147, 0.8423 respectively. Performance did not wane differently in terms of competitive levels and number of institutions, and was able to withstand gradient inference attacks. Discussion HOS-VFL philosophy is centralised performance with minimum communication with retention of raw data local. The method needs to be validated on larger real multi-center datasets before their clinical use.