Alzheimer's Disease (AD) is a complex and prevalent neurodegenerative disorder characterized by distinct brain lesions, including amyloid-β plaques and tau neurofibrillary tangles, which are crucial biomarkers in its diagnosis and progression [14] These lesions vary significantly in their morphological and topographical presentation, contributing to the clinical and pathological heterogeneity of the disease. Current diagnostic frameworks, such as the ABC scoring system, offer standardized methods for assessing AD pathology but are limited in capturing these lesions' nuanced variations and fine-grained heterogeneity [20]. Therefore, advanced methodologies combining digital pathology, artificial intelligence, and spatial morphological analysis have been proposed to refine our understanding of AD and its diverse clinical manifestations [2]. A significant challenge in these methodologies is the identification and quantification of amyloid-β plaques and tau tangles, typically achieved through manual methods or semi-automated proprietary software [1] This process can be labor-intensive and prone to variability due to human observation and expertise. Additionally, the accuracy of these analyses can be affected by variations in histological slide staining. To address these issues, implementing color normalization techniques is essential for standardizing digital pathology workflows and ensuring reliable automated image analysis by leveraging these innovations. Researchers aim to enhance AD pathology's diagnostic precision and understanding, ultimately contributing to improved patient outcomes.

Deep learning methods have made significant progress in this field, providing powerful tools for automatically detecting and segmenting pathological features in AD [18]. These methods offer the potential to overcome the limitations of traditional manual and semi-automated methods by achieving more accurate and reproducible analysis. For instance, convolutional neural networks (CNNs) such as UNet [17] and nnUNet [8] have demonstrated strong performance in medical image segmentation tasks. However, deep learning in AD pathological image segmentation faces two major challenges. The first challenge is the lack of large-scale annotated datasets, which limits the generalization ability of models. Many existing studies rely on small datasets; for example, Wurts et al. [25] used data from a single patient, and Maňoušková et al. [13] used six subjects. In contrast, our study includes data from 15 subjects, providing a more comprehensive dataset for model training and evaluation. The second challenge is the influence of staining variations, which significantly impact deep learning model performance [6]. Stain heterogeneity [11] introduces inconsistencies in model inputs, making segmentation less robust. The underlying mechanisms behind these effects require further investigation, and stain normalization has been recognized as an effective approach to mitigate these variations. To address this, researchers have explored image preprocessing techniques such as color normalization and image enhancement to improve model robustness and accuracy [23], [19], [22].

In this study, we build upon our previous work presented at MICCAI 2022 [9], where we explored the impact of color normalization on deep learning models for neuritic plaque segmentation in whole slide images of human brain tissue. A key advancement in this work is the release of an open-source dataset named as ADNP-15, including both the raw data and images processed with four different stain normalization methods, as well as augmented versions. This significantly extends the MICCAI version, which only provided images processed with two stain normalization methods. Additionally, we establish a comprehensive benchmark by evaluating five widely adopted deep learning models across the four stain normalization techniques, offering deeper insights into their impact on segmentation performance. To further improve segmentation accuracy, we propose a novel image enhancement method that enhances structural details and mitigates staining inconsistencies. Experimental results demonstrate that this enhancement method consistently improves segmentation accuracy, particularly in handling complex tissue structures. All datasets, methods, and code are fully open-source, ensuring transparency and reproducibility while facilitating further advancements in this field.