Nemati Sama, Shabani Hasti, Mahmoudi-Aznaveh Ahmad
Institute of Medical Science and Technology, Shahid Beheshti University, Tehran, Iran.
Cyberspace Research Institute, Shahid Beheshti University, Tehran, Iran.
J Biomed Phys Eng. 2025 Jun 1;15(3):291-298. doi: 10.31661/jbpe.v0i0.2502-1890. eCollection 2025 Jun.
Uneven illumination correction is considered a critical pre-processing step in creating digital images from optical microscopes, particularly in whole-slide imaging (WSI). While deep learning-based methods have suggested new possibilities, they often struggle with generalizing to unseen images and require substantial computational resources. The most common approach for training deep neural networks in this field relies on patch-based processing, which may overlook the global illumination distribution, leading to inconsistencies in correction. This study aimed to identify a key limitation in deep learning models for uneven illumination correction, highlighting the importance of preserving the original image resolution and incorporating a global view of illumination patterns to enhance generalization. To address this, we proposed a new training set design strategy that optimizes neural network performance while utilizing computational resources effectively. Our approach ensures a more uniform correction across entire WSI slides, reducing artifacts and improving image consistency. The proposed strategy enhances model robustness and scalability, making deep learning-based illumination correction more practical for clinical and research applications.
不均匀光照校正被认为是从光学显微镜创建数字图像，特别是在全切片成像（WSI）中的关键预处理步骤。虽然基于深度学习的方法提出了新的可能性，但它们通常难以推广到未见过的图像，并且需要大量的计算资源。该领域中训练深度神经网络最常见的方法依赖于基于补丁的处理，这可能会忽略全局光照分布，导致校正不一致。本研究旨在识别深度学习模型在不均匀光照校正方面的一个关键限制，强调保留原始图像分辨率并纳入光照模式全局视图以增强泛化能力的重要性。为了解决这个问题，我们提出了一种新的训练集设计策略，该策略在有效利用计算资源的同时优化神经网络性能。我们的方法确保在整个WSI切片上进行更均匀的校正，减少伪影并提高图像一致性。所提出的策略增强了模型的鲁棒性和可扩展性，使基于深度学习的光照校正对于临床和研究应用更加实用。
