A bias-resilient client selection analysis for federated brain tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Brain tumor segmentation is challenging due to complex anatomy, image artifacts, and inter-observer variability.
• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise but requires large annotated datasets, posing data privacy concerns.
• Federated learning offers a solution for decentralized model training without centralizing sensitive patient data.

Purpose of the Study:

• To propose and evaluate Fed_WCE_BTD, a novel federated learning framework for brain tumor segmentation.
• To address limitations of traditional federated learning, including institutional biases and inefficient client selection.
• To improve the accuracy and efficiency of automated brain tumor detection and segmentation.

Main Methods:

• Developed Fed_WCE_BTD, combining a modified UNet architecture with federated learning principles.
• Implemented an optimal adaptive client selection strategy to enhance distributed learning performance.
• Validated the proposed method using the BRATS 2021 dataset, considering brain tumor slicing.

Main Results:

• The Fed_WCE_BTD model demonstrated a 1% improvement in detecting enhancing tumors and necrosis compared to existing methods.
• Federated learning significantly improved the dice-coefficient for enhancing tumor segmentation (p<0.05) compared to non-federated approaches.
• Edema identification achieved a comparable dice-coefficient of 80% to the baseline.

Conclusions:

• Fed_WCE_BTD effectively overcomes data privacy barriers in federated learning for brain tumor segmentation.
• The proposed adaptive client selection strategy enhances the performance of decentralized learning.
• This approach offers a viable and accurate solution for computer-aided diagnosis in neuro-oncology.