Han Yechan, Kim Jaeyun, Park Samel, Moon Jong-Seok, Lee Ji-Hey
Department of Medical Science, Soonchunhyang University, Asan, Chungcheongnam-do, South Korea.
Department of AI and Big Data, Soonchunhyang University, Asan, Chungcheongnam-do, South Korea.
Comput Methods Programs Biomed. 2025 Nov;271:109041. doi: 10.1016/j.cmpb.2025.109041. Epub 2025 Aug 24.
Glomeruli are crucial for blood filtration, waste removal, and regulation of essential substances in the body. Traditional methods for detecting glomeruli rely on human interpretation, which can lead to variability. AI techniques have improved this process; however, most studies have used images with fixed magnification. This study proposes a novel magnification-integrated ensemble method to enhance glomerular segmentation accuracy.
Whole-slide images (WSIs) from 12 patients were used for training, two for validation, and one for testing. Patch and mask images were extracted at 256 × 256 size × x2, x3, and x4 magnification levels. Data augmentation techniques, such as RandomResize, RandomCrop, and RandomFlip, were used. The segmentation model underwent 80,000 iterations with a stochastic gradient descent (SGD).
Performance varied with changes in magnification. The models trained on high-magnification images showed significant drops when tested at lower magnifications, and vice versa. The proposed method improved segmentation accuracy across different magnifications, achieving 87.72 mIoU and 93.04 Dice score with the U-Net model.
The magnification-integrated ensemble method significantly enhanced glomeruli segmentation accuracy across varying magnifications, thereby addressing the limitations of fixed magnification models. This approach improves the robustness and reliability of AI-driven diagnostic tools, potentially benefiting various medical imaging applications by ensuring consistent performance despite changes in image magnification.
肾小球对于血液过滤、废物清除以及体内必需物质的调节至关重要。传统的肾小球检测方法依赖人工解读，这可能导致结果的变异性。人工智能技术改进了这一过程；然而，大多数研究使用的是固定放大倍数的图像。本研究提出了一种新颖的集成放大倍数的集成方法，以提高肾小球分割的准确性。
使用12名患者的全切片图像（WSIs）进行训练，2张用于验证，1张用于测试。在256×256尺寸以及x2、x3和x4放大倍数水平下提取图像块和掩码图像。使用了随机缩放、随机裁剪和随机翻转等数据增强技术。分割模型使用随机梯度下降（SGD）进行了80,000次迭代。
性能随放大倍数的变化而变化。在高放大倍数图像上训练的模型在低放大倍数下测试时表现出显著下降，反之亦然。所提出的方法提高了不同放大倍数下的分割准确性，使用U-Net模型实现了87.72的平均交并比（mIoU）和93.04的骰子系数（Dice score）。
集成放大倍数的集成方法显著提高了不同放大倍数下肾小球分割的准确性，从而解决了固定放大倍数模型的局限性。这种方法提高了人工智能驱动的诊断工具的鲁棒性和可靠性，通过确保在图像放大倍数变化时性能一致，可能使各种医学成像应用受益。
