MIE: Magnification-integrated ensemble method for improving glomeruli segmentation in medical imaging.

Area of Science:

• Nephrology
• Computational Pathology
• Medical Imaging Analysis

Background:

• Glomeruli are vital for kidney function, but their detection traditionally relies on subjective human interpretation.
• Existing AI models for glomeruli segmentation often struggle with images of varying magnifications.

Purpose of the Study:

• To develop and evaluate a novel magnification-integrated ensemble method for enhanced glomeruli segmentation.
• To improve the accuracy and robustness of AI-based glomeruli detection across different image magnifications.

Main Methods:

• Whole-slide kidney images were used, with patches extracted at multiple magnification levels (x2, x3, x4).
• Data augmentation techniques were applied to enhance the training dataset.
• A segmentation model, U-Net, was trained using stochastic gradient descent (SGD) with the proposed ensemble method.

Main Results:

• AI model performance significantly degraded when tested at magnifications different from training.
• The magnification-integrated ensemble method demonstrated improved segmentation accuracy.
• The U-Net model achieved 87.72 mIoU and 93.04 Dice score using the proposed method.

Conclusions:

• The proposed magnification-integrated ensemble method effectively enhances glomeruli segmentation accuracy across varying magnifications.
• This approach overcomes limitations of fixed-magnification models, increasing AI diagnostic tool reliability.
• The method offers consistent performance for medical imaging applications, improving diagnostic consistency.