DETECTION OF PATTERNS BETWEEN DIFFERENT STAGES OF ALZHEIMER'S DISEASE USING IMAGE PROCESSING AND STATISTICAL TECHNIQUES
23th INTERNATIONAL İSTANBUL SCIENTIFIC RESEARCH CONGRESS ON LIFE, ENGINEERING, ARCHITECTURE, AND MATHEMATICAL SCIENCES, İstanbul, Türkiye, 20 - 22 Kasım 2025, ss.560-562, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Doi Numarası: 10.30546/19023.978-9952-8605-7-3.2025.0125
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.560-562
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
This study, based on the open-access OASIS database, examined brain structural patterns across different stages of Alzheimer's disease within a multi-layered framework combining image processing and unsupervised/semi-supervised learning. At the core of the study is a proportional indicator defined across 61 individual brain layers and termed “Brain Layer Surface Area” (BLSA); BLSA is calculated as the ratio of pixels belonging to brain tissue in the isolated layer to the total number of isolated area pixels. Of the theoretically expected 32,940 slices from 135 individuals, 31,558 could be analyzed; a robust, voluminous database was established across four health levels (healthy, very mild Alzheimer's, mild Alzheimer's, moderate Alzheimer's), taking into account missing images. The aim is to reveal how layer-based surface area patterns differ across stages and to validate this differentiation with quantitative statistics. The end-to-end process is designed in three stages. In the first stage, raw MRI images were converted to grayscale using OpenCV; noise and small artifacts were cleaned using morphological opening/closing operations. The background was separated using Otsu thresholding, and only the region of interest was preserved using cv2.bitwise_and-based masking. In the second stage, unsupervised segmentation was applied: gray matter, white matter, and void components were obtained in each layer using K-Means (n_clusters=3); then, the components were reduced to “tissue” and “void” superclusters using K-Means (n_clusters=2) again. In this step, to strengthen the similarity between images, the intermediate layer representation of VGG16 (7 × 7 × 512 = 25,088-dimensional vector) was utilized, and it was determined which supercluster the components resembled more based on Euclidean distances. In the third stage, BKYA values were calculated for all layers, serially normalized on a layer-by-layer basis, and standardized into a four-phase × 61-layer format for statistical modeling. Data quality and distribution characteristics have been verified using statistical tests. In the preliminary analysis, the extreme values (0 and 100) of the BKYA distributions were removed; Kolmogorov-Smirnov tests indicated that the parametric assumptions were weak and that non-parametric methods would be more appropriate. Basic statistics quantitatively confirmed the pattern evolving according to stages. The mean BKYA was 89.21% in the healthy group, 85.87% in the very mild group, 86.60% in the mild group, and 79.90% in the moderate group. The coefficient of variation (CV) was found to be highest, at ~17%, particularly in the moderate group. This indicates increased inter-individual heterogeneity in advanced stages. Skewness and kurtosis statistics suggest that the distributions are mostly flat, but that the shape may show more pronounced asymmetry in advanced stages. The relationship between stages in terms of proximity/distance was examined using Spearman's rho coefficients. The lowest correlation is between healthy and moderate stages (ρ≈0.731); this reflects the expected structural differentiation. The very mild and mild stages, on the other hand, showed high similarity (ρ≈0.918). This finding supports the expectation that structural differentiation in the early stages develops in a gradual and overlapping continuum, consistent with the expected structure. The relationship between healthy and very mild stages was also relatively high (ρ≈0.912). This indicates that structures are still largely preserved in the stage immediately following healthy, but leakage-like changes have begun in selected layers. To test the statistical significance of differences between groups, Friedman's S test was applied, yielding a test statistic of approximately 114.74 under the χ² approach; this value significantly exceeds the table value (7.815) for 3 degrees of freedom, providing strong evidence of a general difference at the p<0.05 level. In the post-hoc stage, the Wilcoxon Signed-Rank Test (with multiple comparison correction) was used. According to this, the healthy group and all other Alzheimer's stages, as well as the moderate stage and all other stages, are significantly differentiated. No statistically significant difference was found between the very mild and mild stages (p=0.134). This result, combined with the correlation findings, indicates the appropriateness of considering early stages as a “process band” and that discriminatory power can be enhanced, particularly with layer-selective criteria. The cluster structure of spatial patterns was also examined in the stage-layer plane using DBSCAN. This noise-sensitive, density-based approach reveals natural clusters and outliers without requiring labels. Maps obtained with hyperparameters selected through optimization show that BKYA ratios exhibit a marked downward trend as the stage progresses, particularly in the first 30 layers; in subsequent layers (approximately 30–61), inter-stage values converge more closely. This two-part dynamic suggests that cortical surface area loss is more pronounced in the early depths and enters a plateau trend in the advanced layers. Methodologically, this thesis goes beyond approaches commonly found in the literature that focus on specific sub-structures (such as the hippocampus, selected cortical areas, etc.), modeling the entire brain layer-by-layer and through unlabeled segmented structures. The combined use of two-stage K-Means + VGG16 representation validated the brain tissue-void distinction not only by density but also by content similarity; the BKYA variable emerged as an operational biomarker capable of tracking stage progression. The findings simultaneously support both strong statistical evidence of interclass differences and the idea of continuity between early stages. As a result, the developed workflow (isolation, unsupervised segmentation, deep representation-based clustering, BKYA calculation, nonparametric statistics, and density-based clustering) can reliably capture layer-sensitive structural differentiation independently of clinical labels. The stage sensitivity of BKYA and its high discriminative power in early layers indicate the method's applicability in screening/pre-diagnostic scenarios. Future studies should combine BKYA with longitudinal data for intra-individual tracking, 3D volumetric measures (e.g., cortical thickness, surface roughness), and strengthen generalizability through multicenter external validation. This holistic approach demonstrates the potential for integrating image-based, label-free, and layer-sensitive analytics into clinical decision support systems for Alzheimer's disease.