High-Precision Detection of Cells and Amyloid-β Using Multi-Frame Brightfield Imaging and Quantitative Analysis

Abstract

This study presents a novel method for high-precision detection and quantitative evaluation of the spatial relationship between cells and amyloid-$\beta$ (A$\beta$) in time-lapse brightfield microscopy images. Achieving accurate detection of non-fluorescent cells and A$\beta$ deposits requires high-quality video images free from noise, distortion, and frame-to-frame luminance flicker. To this end, we employ a robust preprocessing pipeline that combines multi-frame integration with vignetting correction to enhance image quality and reduce luminance variability across frames. Key preprocessing steps include background correction via two-dimensional polynomial fitting, temporal smoothing of luminance fluctuations, histogram matching for luminance normalization, and dust artifact removal based on intensity thresholds. This enhanced imaging approach enables accurate identification of A$\beta$ aggregates, which typically appear as jelly-like structures and are difficult to detect under standard brightfield conditions. Furthermore, we introduce a quantitative index to assess the spatial relationship between cells and A$\beta$ concentrations, facilitating detailed analysis under varying A$\beta$ levels.

1. Introduction

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that typically manifests in later adulthood. Among the prevailing theories of its pathogenesis, the amyloid hypothesis has gained widespread acceptance. This hypothesis posits that amyloid-beta (A$\beta$) peptides play a central role in neuronal degeneration by inducing cytotoxicity and triggering inflammatory responses in the brain [1]. In particular, A$\beta$ aggregates are believed to exacerbate tissue damage, making them a major focus of AD research. Accurate detection of A$\beta$ aggregates remains challenging due to their low contrast and gel-like distribution in time-lapse brightfield microscopy images. Additional complications arise from background noise, uneven illumination, and vignetting artifacts commonly present in such imaging sequences. Therefore, conventional visualization methods often rely on subjective evaluations using image analysis tools such as ImageJ (version 1.54p) [2]. In this study, a method is proposed for high-precision detection and quantitative evaluation of the spatial relationship between cells and A$\beta$ aggregates in time-lapse brightfield microscopy images. The method incorporates a comprehensive image preprocessing pipeline consisting of vignetting correction, multi-frame image enhancement, background normalization, and noise reduction. These refinements enable accurate identification of subtle A$\beta$ features that are typically difficult to detect. To quantitatively assess A$\beta$ aggregation near cells, four evaluation metrics have been developed, focusing on the amount and spatial characteristics of aggregation. Validation was conducted using time-lapse images of human neuroblastoma SH-SY5Y cells exposed to A$\beta$ at concentrations ranging from 0.625 μM to 20 μM. Effectiveness of the proposed method was confirmed through analysis of dynamic changes in cell–A$\beta$ interactions. The structure of this paper is as follows: Section 2 reviews related work and contextualizes the proposed approach within the literature. Section 3 describes the preprocessing procedures for enhancing A$\beta$ visibility. Section 4 defines the quantitative evaluation metrics. Section 5 presents the validation experiments. Section 6 provides analysis of the results, and Section 7 concludes the study.

2. Related Works

The application of artificial intelligence (AI) technologies, particularly deep learning models, to cellular image analysis has advanced significantly in recent years. A wide range of methods have been proposed for automated cell detection and segmentation, demonstrating notable performance improvements. Among these, Cellpose(version 3.0) [3] has emerged as a leading approach. Built upon the U-Net architecture, a well-established convolutional neural network (CNN) model, Cellpose learns a vector field that represents the directional “flow” of cells. This enables precise segmentation and detection of individual cells, including those with complex morphologies or those present in densely clustered arrangements [3].

In parallel with cell detection, significant progress has been made in cell tracking, which involves linking individual cells across time points in time-lapse microscopy. This technique enables quantitative analysis of various dynamic cellular behaviors, such as migration, proliferation, division, fusion, and apoptosis [4,5]. These studies rely on accurately segmented cells from each frame, often using deep learning-based approaches such as Cellpose as the foundation for temporal tracking.

In contrast, relatively little research has addressed the detection of A$\beta$ aggregates in time-lapse brightfield microscopy using AI techniques. Prior investigations of A$\beta$ detection have primarily focused on large-scale accumulation visualized through PET and MRI imaging of the human brain [6,7], rather than cellular-scale detection in optical microscopy.

Two main challenges have limited the development of AI-based A$\beta$ detection methods in brightfield microscopy:

• Difficulty in Direct Visualization: A$\beta$ aggregates, whether intra- or extracellular, exhibit extremely low contrast in brightfield images, making them inherently difficult to visualize and distinguish from the background.
• Lack of Distinct Morphological Features: Unlike cells, A$\beta$ aggregates do not display clear or consistent morphological patterns. This absence of distinctive visual features complicates the training of AI models, as the features necessary for accurate detection are often ambiguous or absent in brightfield modalities.

• In this study, we proposed a detection method of A$\beta$ in time-lapse brightfield microscopy 72 that does not utilize deep learning models.

3. Preprocessing for Time-Lapse Brightfield Microscopy

Brightfield microscopy imaging of A$\beta$ and cells presents several inherent challenges. A$\beta$ appears as a translucent, gel-like structure without well-defined edges, while cells often exhibit weak contrast and diffuse boundaries. This ambiguity, combined with imaging artifacts such as vignetting and flickering, makes it difficult to detect and quantify targets with high confidence. To address these limitations, we designed a two-part preprocessing pipeline for A$\beta$ and cell image enhancement, respectively.

3.1. Challenges in A$\beta$ Detection

Brightfield microscopy imaging of A$\beta$ presents several inherent challenges. Unlike fluorescent labels, A$\beta$ appears as a translucent, gel-like structure without well-defined edges. This ambiguity, combined with a series of imaging artifacts, makes it difficult to detect and quantify A$\beta$ with high confidence. Specifically, we identified three major issues in the raw image data.

• Vignetting artifacts: Radial brightness decay caused by lens flare and uneven illumination, which obscure object boundaries. An example of this effect is shown in Figure 1.
• Temporal flickering: Frame-to-frame brightness fluctuations due to unstable illumination, unrelated to biological signals.
• Small bright debris: Foreign particles that mimic A$\beta$ aggregates, leading to incorrect detection if unfiltered.

Without correcting these issues, A$\beta$ segmentation becomes highly unstable, especially in low-concentration conditions, and may distort downstream analyses of its accumulation behavior and interaction with cells. Therefore, we developed a comprehensive preprocessing pipeline aimed at improving brightness consistency, suppressing noise, and enhancing A$\beta$ visibility before segmentation.

3.2. A$\beta$ Image Preprocessing

Therefore, we implemented the following preprocessing steps to address these issues. The preprocessing framework consists of the following five stages, each targeting a specific source of noise or instability.

(1)

Polynomial-Based Background Correction

To address vignetting and halo effects, we modeled the background brightness distribution using a two-dimensional polynomial surface:

$$P(x,y)=\sum _{i=0}^n\sum _{j=0}^{n-i}{a}_{ij}{x}^i{y}^j$$

where $(x,y)$ represents the pixel coordinate and $a_{ij}$ are the polynomial coefficients. A third-order polynomial ($n=3$) was typically used. Each frame was divided by its normalized background model to suppress asymmetric brightness decay and enhance the visibility of A$\beta$ structures [8].

To quantitatively evaluate the background fitting quality, we computed the absolute residuals between the original image and the fitted background surfaces using first- to fourth-order polynomials. Figure 2 shows the histograms of absolute residuals for each polynomial order. The mean absolute errors (MAE) were 0.927, 0.495, 0.482, and 0.497 for degrees 1 to 4, respectively.

MAE reflects the average pixel-wise deviation between the original and fitted background intensities. A lower MAE indicates a better approximation of the underlying background illumination [9]. The significant drop from degree 1 to 2 suggests that first-order (planar) fitting fails to capture the image’s radial gradient, while higher-order models more accurately represent the vignetting pattern. Among these, degree 3 achieved the lowest MAE, with only marginal differences compared to degrees 2 and 4. Considering both performance and model simplicity, we selected third-order polynomials as the default background fitting method.

To further evaluate the generalizability of our approach, we conducted a comparative analysis with alternative correction methods, as described in Section 6.

(2)

Temporal Brightness Normalization

We computed the 10th percentile pixel intensity for each frame and fitted a regression curve over time. Each frame was scaled accordingly to normalize illumination drift across the sequence.

(3)

Flicker Smoothing via Exponential Moving Average (EWMA)

To suppress short-term brightness fluctuations, EWMA smoothing was applied with a decay factor $\alpha =0.6$, preserving gradual trends while eliminating abrupt changes [10].

(4)

Tone Stabilization through Percentile Scaling and Histogram Matching

Each frame was normalized to the 1st–99th percentile intensity range, followed by histogram matching to a reference frame to ensure tonal consistency across time [11].

(5)

Artifact and Debris Removal via Brightness Profiling and Inpainting

Bright spots from debris or noise were detected using intensity line profiles and removed via OpenCV’s inpaint function (Figure 3).

(6)

Comparison with Conventional Background Correction Methods

To further validate the effectiveness of our polynomial-based correction strategy, we conducted a comparative analysis with six widely used background enhancement methods. These include histogram-based, statistical, model-based, and morphological techniques:

• Adaptive Histogram Equalization (AHE): Enhances local contrast by redistributing pixel intensities within small regions. While effective in improving local details, it is prone to noise amplification and over-enhancement in homogeneous areas [12].
• Contrast-Limited Adaptive Histogram Equalization (CLAHE): An extension of AHE that limits histogram amplification to reduce noise. Commonly used in biomedical imaging to enhance tissue visibility while avoiding over-saturation [13].
• CV Map (Coefficient of Variation): Computes the local standard deviation-to-mean ratio to highlight regions of high intensity variation. Particularly useful in heteroscedastic image preprocessing and noise-insensitive enhancement [14,15].

Figure 4 presents the results of applying these methods to the same representative A$\beta$ frame. Polynomial fitting (degree 3) provides a balance between structural preservation and background suppression, yielding visually consistent enhancement across varying regions.

In contrast to A$\beta$, cell detection relies on texture and local contrast. We therefore adopted variance-based filtering to emphasize cellular structures.

3.3. Cell Image Preprocessing

To ensure accurate segmentation of cells from brightfield microscopy images, we applied a preprocessing pipeline specifically designed to enhance weak cellular boundaries and suppress non-cellular background artifacts. The key strategy was to utilize local temporal variance to distinguish between dynamic foreground (cells) and static background.

Local Variance Filtering Based on Temporal Brightness Fluctuations

Due to subtle motion and morphological changes, live cells in brightfield sequences exhibit slight but consistent frame-to-frame brightness variations. In contrast, the background remains largely static over time. We leveraged this property by computing the pixel-wise variance across a short temporal window (e.g., 5 consecutive frames). This operation was performed in a block-wise manner to suppress noise and capture regional intensity dynamics [16].

Regions with high variance were interpreted as likely cell areas, since they reflect localized brightness changes due to cell movement or shape fluctuation. In contrast, regions with low variance were treated as static background. The resulting variance map was then resized to the original resolution via bilinear interpolation, preserving spatial precision. As shown in Figure 5, background correction improves the visibility of cell structures.

While the current approach relies on local variance filtering due to its simplicity, parameter-free nature, and robustness to subtle cell motion, we acknowledge that deep learning-based segmentation methods such as Cellpose [3] offer higher accuracy and flexibility. Incorporating such methods is part of our future work, particularly for evaluating segmentation performance under more complex conditions.

3.4. Summary

This preprocessing pipeline systematically corrects for optical and temporal artifacts common in brightfield microscopy. For A$\beta$, the method improves visibility by removing vignetting, flicker, and debris. For cells, local variance filtering improves structural clarity and segmentation accuracy. These enhancements collectively facilitate reliable downstream analysis without the need for staining or advanced microscopy techniques.

4. Detection of A$\mathit{\beta }$ and Cells, and Quantification of A$\mathit{\beta }$ Aggregation Around Cells

In this section, we present a comprehensive framework for detecting both A$\beta$ aggregates and cells in time-lapse brightfield microscopy images, with the goal of quantifying the spatial accumulation of A$\beta$ around cells. Building on the preprocessing pipeline described in the previous section, we introduce detection methods tailored to the distinct structural and brightness characteristics of A$\beta$ and SH-SY5Y cells. To facilitate objective evaluation of A$\beta$ aggregation in proximity to cells, we define four quantitative metrics that capture various aspects of overlap and spatial association between A$\beta$ and cellular regions. These metrics enable detailed comparisons across different experimental conditions and A$\beta$ concentrations, providing a robust basis for interpreting A$\beta$–cell interactions in Alzheimer’s disease research. The following sections provide detailed descriptions of the detection methods applied to A$\beta$ aggregates and SH-SY5Y cells.

4.1. Segmentation Methodology

To quantitatively evaluate the spatial relationship between A$\beta$ aggregates and cells, it is essential to obtain accurate binary masks for both. Owing to their differing structural and optical properties, separate segmentation approaches were employed for A$\beta$ and cells.

4.1.1. A$\beta$ Segmentation

A$\beta$ aggregates exhibit gel-like translucency and heterogeneous morphologies that vary with concentration and aggregation stage. To accurately segment both sparse and dense aggregates, we developed a hybrid strategy tailored to concentration-dependent patterns.

• For low-concentration samples (≤5 μM): In these cases, A$\beta$ aggregates are sparse and typically appear as small, isolated structures. After background correction, the intensity distribution becomes sufficiently distinct to allow for global thresholding. We empirically selected a fixed threshold for each concentration group based on representative brightness histograms. This approach minimizes false positives and ensures high specificity under relatively uniform background conditions.
• For high-concentration samples (≥10 μM): At higher concentrations, aggregation is more widespread, with considerable variability in brightness and density across time and space. Therefore, we employed a semi-automatic, dynamic thresholding strategy:
  • Step 1: Manual inspection was performed on several keyframes (e.g., frames 0, 30, and 60), and optimal threshold values were empirically determined for each.
  • Step 2: Threshold values for intermediate frames were obtained via linear interpolation. To suppress abrupt transitions caused by local brightness anomalies, the frame-to-frame change in threshold was constrained to ≤0.02.
  • Step 3: A binary mask was generated for each frame using the interpolated threshold.
  • Step 4: To enhance temporal consistency and reduce transient noise, a five-frame temporal voting filter was applied: a pixel is retained as foreground (A$\beta$) only if it is labeled as foreground in at least 3 of the 5 adjacent frames (two preceding, current, and two following). This majority-vote filter helps preserve persistent structures while discarding flickering artifacts.
  • Step 5: To further refine the segmentation result, we removed small objects with fewer than 500 pixels (which are typically noise or debris) and applied a morphological closing operation to connect fragmented A$\beta$ clusters, filling small gaps and smoothing boundaries.

The classification of ≤5 μM as low and ≥10 μM as high concentration is supported by recent studies on A$\beta$ supersaturation and nucleation theory. According to a thermodynamic phase diagram proposed in a recent study [17], concentrations below 5 μM fall within the metastable zone, where A$\beta$ tends to remain unaggregated or transiently oligomerized. In contrast, concentrations ≥10 μM fall into the nucleation zone, where rapid aggregation and fibril formation dominate. This division enables more accurate segmentation strategies tailored to distinct aggregation behaviors.

To assess the spatial relationship between A$\beta$ accumulation and cellular regions, it is essential to also detect cell boundaries with high accuracy. The following section describes our approach for segmenting SH-SY5Y cells from the same time-lapse brightfield microscopy images.

4.1.2. Cell Segmentation

As described in Section 3, the cell preprocessing step leverages local variance to highlight pixel regions with fluctuating brightness due to cell movement. This enhances contrast between the dynamic foreground (cells) and the relatively static background.

To segment cell regions, we applied a confidence-based method.

• Step 1: A block-wise local variance map was computed to enhance structures with temporal brightness changes—mainly corresponding to moving or morphologically active cells.
• Step 2: The variance map was transformed into a confidence map using a sigmoid activation function. This step assigns higher confidence to regions more likely to be cells [18].
• Step 3: A fixed threshold was applied to the confidence map to obtain a binary cell mask.
• Step 4: Small connected components were removed to eliminate debris or noise.

With both A$\beta$ and cell regions successfully segmented, we proceeded to analyze their spatial relationships across time-lapse sequences. To quantify the degree and pattern of A$\beta$ accumulation around cells, we introduce a set of four metrics that provide complementary insights into overlap extent, coverage, and aggregation behavior under different experimental conditions.

4.2. Metrics for A$\beta$–Cell Relationship Analysis

To systematically evaluate how A$\beta$ interacts with cells over time, we propose a set of four interpretable metrics that jointly capture spatial co-localization, aggregation dynamics, and relative coverage. While conventional studies often focus on global A$\beta$ quantities or qualitative inspection, our metrics aim to quantify the heterogeneous distribution and accumulation behavior of A$\beta$ at the cellular level.

These metrics were computed for each frame and concentration condition using the binarized masks of A$\beta$ and cells obtained from the preprocessing pipeline.

• Average Number of Overlapping Pixels per Cell: This metric calculates the mean number of pixels where A$\beta$ overlaps with each detected cell at each time point. It provides a direct measure of per-cell A$\beta$ burden and is sensitive to how A$\beta$ selectively adheres to or accumulates on certain cells. The relationship between A$\beta$ aggregates and individual cells is visually illustrated in Figure 6.
• Overlap Ratio Relative to A$\beta$ Area: This metric calculates the proportion of A$\beta$ pixels that are co-localized with cell regions. A high value implies preferential A$\beta$ clustering near cells, while a low value suggests diffuse or cell-independent aggregation. This relationship is visually illustrated in Figure 7.
• Overlap Ratio Relative to Cell Area: This indicator computes the ratio between the overlapping pixels and the total cell area. It reflects how much of each cell is physically covered by A$\beta$, accounting for size variations across different cell morphologies. This relationship is visually illustrated in Figure 8.
• Number of A$\beta$ Aggregates: We count the number of distinct A$\beta$ clusters per frame, defined as connected components above a minimum size threshold (excluding artifacts). This metric reflects overall aggregation dynamics and how they vary with concentration and over time. This relationship is visually illustrated in Figure 9.

These metrics provide a multi-faceted and biologically meaningful view of A$\beta$–cell interactions, enabling downstream comparison across different concentration conditions and time windows.

5. Experimental Results

To validate the segmentation and quantification methods proposed in Section 3 and Section 4, we conducted a series of time-lapse imaging experiments using human neuroblastoma SH-SY5Y cells treated with varying concentrations of A$\beta$. The objective was to assess whether the preprocessing and detection strategies developed in this study enable robust identification of A$\beta$ aggregates and their spatial relationships with cells under biologically relevant conditions. By applying the four quantitative metrics introduced earlier, we further examined how A$\beta$ accumulation patterns vary with concentration and exposure duration. The following subsections describe the experimental setup in detail.

5.1. Experimental Setup

The cells used in this study were human neuroblastoma SH-SY5Y cells. Time-lapse imaging was conducted under controlled conditions while varying the A$\beta$ concentration. The experimental procedures are summarized below (Box 1).

Box 1. Imaging Procedure.

• Day 1: Plate Coating

IWAKI 96-well glass-bottom plates (5866–096, IWAKI, Shizuoka, Japan) were coated with 20 μg/mL fibronectin (Gibco, Thermo Fisher Scientific, Waltham, MA, USA; Cat. No. 33016-015) solution and incubated overnight at 37 °C with 5% CO₂ to enhance cell adhesion to the plate surface.

• Day 2: Cell Seeding

After washing the wells with PBS (137 mmol/L NaCl, 8.1 mmol/L Na₂HPO₄, 2.68 mmol/L KCl, 1.47 mmol/L KH₂PO₄; self-prepared) and DMEM (Wako, Odawara, Japan), SH-SY5Y human neuroblastoma cells (KAC Co., Ltd., Kyoto, Japan) were seeded at $0.4\times {10}^4$ cells per well in 200 μL of DMEM medium. The plates were incubated overnight at 37 °C with 5% CO₂.

• Day 3: A$\beta$ Treatment

A$\beta$ solutions were prepared using Human A$\beta$₄₂ peptide (4349-v, Peptide Institute Inc., Osaka, Japan) and Cys-conjugated A$\beta$₄₀ (23519, Anaspec Inc., Fremont, CA, USA) in DMEM/F12 (1:1) medium (Gibco, Life Technologies, Waltham, MA, USA). To enable fluorescence-based validation and visualization, Qdot-labeled A$\beta$ (Qdot^TM 605 ITK^TM Amino (PEG) Quantum dot; Q21501MP, Thermo Fisher Scientific, Waltham, MA, USA) was used in combination with Cys-conjugated A$\beta$₄₀, as described previously [19]. Each well was treated with 100 μL of the respective solution and incubated for 24 h under the same temperature and CO₂ conditions. Previous studies have shown that higher A$\beta$ concentrations lead to more pronounced aggregation and increased interaction with cells [20]

• Day 4: Microscopy Observation

Time-lapse imaging was conducted using an inverted microscope (Ti-E, Nikon, Tokyo, Japan) equipped with a PlanApo $\lambda$ 20 × /0.75 NA objective lens (Nikon), a color CMOS camera (DS-Ri2, Nikon, Tokyo, Japan), and a TRITC filter set (TRITC-A-Basic-NTE, excitation: 532–552 nm, emission: 594–646 nm). During observation, cells were maintained in a chamber at 37 °C with 5% CO₂ (INUBTF-WSKM; Tokai Hit, Fujinomiya, Japan). Images were acquired and analyzed using NIS-Elements AR software (Nikon, Tokyo, Japan).

5.2. Segmentation Results

To validate the effectiveness of the segmentation methods introduced in Section 4, we applied the proposed pipelines to the acquired time-lapse brightfield microscopy images under varying A$\beta$ concentrations. Representative examples of A$\beta$ and cell segmentation outcomes are presented below.

Figure 10 illustrates the impact of postprocessing steps on the binarized detection of A$\beta$ aggregates. For cell segmentation, Figure 11 and Figure 12 demonstrate how the variance-based method combined with sigmoid activation and small-object removal enables clear delineation of individual cell regions. These visual results support the robustness of the proposed approach in handling diverse cellular morphologies and A$\beta$ distribution patterns in brightfield images.

This approach exploits the dynamic nature of cells in time-lapse brightfield imaging and does not require fluorescence labeling. It is robust to variations in shape, contrast, and background noise.

To further support the reliability of the segmentation results, we analyzed the temporal trends of the segmented areas across time-lapse sequences.

Validation via Area-Time Curves

Due to the absence of ground-truth masks, we indirectly assessed segmentation quality by analyzing the temporal variation of detected areas (see Figure 13). Smooth upward trends in high-concentration conditions and low-noise flat curves at low concentrations support the validity of our segmentation pipeline.

To indirectly assess the temporal stability of A$\beta$ segmentation in the absence of ground truth annotations, we plotted the total segmented A$\beta$ area for each frame over time. Representative results (Figure 13) reveal the following trends:

• At high concentrations (e.g., 20 μM), the segmented A$\beta$ area increases over time and then stabilizes, matching the expected tendency of A$\beta$ to accumulate progressively under such conditions.
• At low concentrations (e.g., 2.5 μM), the segmented area remains low and exhibits irregular fluctuations, likely due to weak A$\beta$ presence and sensitivity to noise.

These trends suggest that the segmentation method behaves stably in high-signal scenarios and avoids over-segmentation in low-signal frames.

Building upon the validated segmentation outputs, we next focus on quantifying how A$\beta$ aggregates spatially relate to cells across different experimental conditions.

5.3. Quantitative Indicators of A$\beta$–Cell Interaction

To investigate how A$\beta$ interacts with cells under different conditions, we designed four quantitative indicators:

• Average Overlapping Pixels per Cell: measures how much A$\beta$ contacts individual cells.
• Overlap Ratio Relative to A$\beta$ Area: evaluates how concentrated A$\beta$ is around cell regions.
• Overlap Ratio Relative to Cell Area: captures the percentage of each cell’s surface that is in contact with A$\beta$.
• Number of A$\beta$ Aggregates: counts cluster structures to estimate aggregation extent.

These indicators were computed for all time points and concentration groups to reveal dose- and time-dependent trends.

5.4. Temporal Trends and Concentration-Dependent Behavior

5.4.1. Average Overlapping Pixels per Cell

Figure 14 demonstrates that at high concentrations (10–20 μM), the number of overlapping pixels per cell increases significantly over time, reflecting sustained A$\beta$ accumulation on cell surfaces. Lower concentrations show stable or noisy patterns, lacking strong aggregation signals.

5.4.2. Overlap Ratio Relative to A$\beta$ Area

As shown in Figure 15, cells exposed to high A$\beta$ concentrations gradually acquire coverage ratios of 30–60%, while low concentrations fluctuate unpredictably. This confirms selective membrane binding at elevated A$\beta$ levels.

5.4.3. Overlap Ratio Relative to Cell Area

Figure 16 reveals that at 20 μM, over 40% of total A$\beta$ localizes around cell surfaces. This cellular localization decreases with concentration, supporting a saturation-like interaction mechanism.

5.4.4. Number of A$\beta$ Aggregates

Figure 17 illustrates that A$\beta$ aggregates are more frequent and widespread under high-concentration conditions, where mean overlap per cell can reach 40%. This implies non-uniform attachment, possibly modulated by membrane characteristics.

The simulation results confirm that the proposed segmentation pipeline enables stable and biologically meaningful detection of both A$\beta$ and cells. The four proposed indicators effectively captured dose-dependent interaction patterns and aggregation dynamics, supporting the A$\beta$ toxicity hypothesis in Alzheimer’s pathology.

6. Analysis

To elucidate the interaction between A$\beta$ and SH-SY5Y cells under varying concentrations, we quantitatively analyzed time-lapse microscopy data using four proposed indicators: (1) average overlapping pixels per cell, (2) overlap ratio relative to cell area, (3) overlap ratio relative to A$\beta$ area, and (4) number of A$\beta$ aggregates. The segmentation pipeline described earlier enabled robust mask generation for both A$\beta$ and cells, providing a reliable basis for downstream analysis.

6.1. Dose-Dependent A$\beta$ Accumulation on Cell Surfaces

Figure 14 presents the average number of overlapping pixels per cell over time. At high concentrations (10–20 μM), A$\beta$ accumulation increases steadily throughout the observation window, indicating progressive membrane attachment. By contrast, in low-concentration conditions (≤5 μM), the overlap remains low and fluctuates irregularly, with no sustained upward trend. This supports a threshold-like behavior in A$\beta$ interaction, wherein significant adhesion occurs only above a certain concentration level.

6.2. Membrane Binding Efficiency and Selectivity

The overlap ratio relative to the A$\beta$ area (Figure 15) complements this perspective by quantifying how much of the total A$\beta$ is localized near cells. In the 20 μM group, this ratio reaches 0.4–0.6 by the end of the observation period, suggesting that a large portion of A$\beta$ preferentially accumulates at cell surfaces. In contrast, at lower concentrations, the ratio fluctuates more irregularly, likely due to sparse aggregation and susceptibility to noise. Despite occasional spikes, the high-concentration groups show a consistent upward trend indicative of biologically plausible cell-targeting behavior.

The overlap ratio relative to cell area (Figure 16) provides insight into how much of each cell’s surface is covered by A$\beta$. At 20 μM, many cells show 30–60% coverage, suggesting substantial membrane coating. Meanwhile, lower concentrations display unstable or negligible coverage, consistent with minimal binding. This suggests that efficient membrane coating occurs only when sufficient A$\beta$ is present to overcome the interaction threshold.

6.3. Heterogeneity in Aggregation Dynamics

Figure 17 shows the mean overlap per cell, revealing that A$\beta$ does not bind uniformly. Even at high concentrations, per-cell values vary, implying heterogeneity in local susceptibility, potentially driven by differences in membrane composition or microenvironment. This index, when combined with the others, highlights the nuanced, non-uniform nature of A$\beta$ accumulation—important for understanding differential cytotoxicity responses in neural populations.

6.4. Implications for the A$\beta$ Toxicity Hypothesis

Together, these indicators support the amyloid hypothesis, which posits that A$\beta$ aggregation induces neurotoxicity once a critical threshold is surpassed [20,21]. Our results suggest the following:

• A$\beta$ surface accumulation initiates when concentrations exceed 5–10 μM.
• Aggregation accelerates beyond frame 50 (approx. 12 h), indicating a transition from oligomer formation to extensive clustering.
• Cell surface coverage by A$\beta$ surpasses 30% under 20 μM conditions, a level reported in prior studies to impair membrane integrity and receptor function [22].

The temporal patterns across all four metrics reveal a clear dose-response relationship and dynamic progression of toxicity. Notably, the pipeline successfully captured early aggregation phases and late-stage saturation using label-free brightfield images, overcoming prior reliance on fluorescent markers.

6.5. Limitations and Future Directions

Despite the overall robustness of our findings, several limitations should be acknowledged. Minor frame-to-frame fluctuations may be influenced by illumination shifts or residual flicker, though mitigated via temporal filtering. Additionally, the current analysis is based on population-level metrics rather than single-cell tracking, which may overlook cell-specific behaviors and responses.

Moreover, while this study primarily focuses on preprocessing and detection of A$\beta$ aggregates, the segmentation of cells was performed using a simple local variance-based method to extract overall cellular regions. This approach is effective for capturing general spatial relationships but lacks precision in delineating individual cells and their morphological dynamics. Furthermore, the current quantitative indicators primarily focus on morphological features, such as spatial overlap or coverage ratios, without directly capturing the physiological functional states of individual cells. While these morphological measures provide indirect insights into potential cell responses (e.g., clustering or shrinkage), they may not fully reflect functional consequences such as apoptosis, necrosis, or viability loss caused by A$\beta$ exposure.

Future extensions of this framework may include:

• Incorporating deep learning-based segmentation tools such as Cellpose [3] to enable accurate instance segmentation of individual cells in brightfield images;
• Analyzing temporal changes in cell activity, including protrusion formation, shrinkage, and motility under different A$\beta$ concentrations;
• Implementing single-cell tracking to monitor dynamic behaviors over time and quantify responses to A$\beta$ exposure at the individual-cell level;
• Quantifying morphological indicators such as circularity or aspect ratio to capture structural alterations related to cell stress or death;
• Estimating instantaneous velocities to assess changes in cell migration dynamics in response to A$\beta$ accumulation.

7. Conclusions

This study focused on the quantitative evaluation of A$\beta$–cell interactions in time-lapse brightfield microscopy images. Original image preprocessing was proposed to enable accurate identification of both cells and A$\beta$ aggregates under challenging imaging conditions. In addition, four quantitative metrics were developed to assess the degree and spatial characteristics of A$\beta$ aggregation around cells.

Validation experiments using human neuroblastoma SH-SY5Y cells exposed to A$\beta$ concentrations ranging from 0.625 μM to 20 μM demonstrated the method’s ability to capture temporal changes in A$\beta$ aggregation with high precision. The results confirmed the effectiveness of the proposed approach in extracting meaningful quantitative information from brightfield time-lapse sequences.

Given its scalability and compatibility with label-free imaging, the proposed method represents a promising tool for high-throughput drug screening applications and mechanistic studies of A$\beta$ dynamics in neurodegenerative disease research.