Multimodal OF-MTMFL: A Semi-Supervised Mean Teacher Model for Histopathological Image Segmentation.

Abstract

In the rapidly advancing field of histopathological image analysis, accurate segmentation of critical features is crucial for medical diagnostics, as it enables pathologists to make precise decisions. The proposed One Former-based Mean Teacher Model with Federated Learning (OF-MTMFL) system combines cutting-edge semi-supervised learning and federated learning techniques to tackle issues such as limited annotated data and class imbalance. The framework utilizes a mean teacher architecture, where the student model, guided by a focal loss function, prioritizes high-confidence regions in unlabeled data, while the teacher model ensures consistency through Exponential Moving Average (EMA) updates. To further enhance segmentation accuracy, multi-scale attention modules are employed for robust feature extraction. Additionally, the system incorporates a Federated Learning mechanism that allows multiple institutions to collaborate without sharing raw data, including datasets from the Cancer Genome Atlas (TCGA). The results from the analysis of the TCGA dataset indicate that the proposed OF-MTMFL model achieved mean concordance index (c-index) scores of 0.700 ± 0.030 for Bladder Urothelial Carcinoma (BLCA), 0.720 ± 0.040 for Breast Invasive Carcinoma (BRCA), 0.860 ± 0.025 for Glioblastoma & Lower Grade Glioma (GBMLGG), 0.690 ± 0.035 for Lung Adenocarcinoma (LUAD), and 0.740 ± 0.045 for Uterine Corpus Endometrial Carcinoma (UCEC). The overall performance score of the OF-MTMFL model across these cancer types is 0.740, demonstrating particularly strong results in GBMLGG while maintaining competitive scores in the other cancer types. The standard deviations reported reflect the variability of the model's performance across different samples within each category.