Correlating Histopathological Microscopic Images of Creutzfeldt–Jakob Disease with Clinical Typology Using Graph Theory and Artificial Intelligence

Abstract

Creutzfeldt–Jakob disease (CJD) is a rare, degenerative, and fatal brain disorder caused by abnormal proteins called prions. This research introduces a novel approach combining AI and graph theory to analyze histopathological microscopic images of brain tissues affected by CJD. The detection and quantification of spongiosis, characterized by the presence of vacuoles in the brain tissue, plays a crucial role in aiding the accurate diagnosis of CJD. The proposed methodology employs image processing techniques to identify these pathological features in high-resolution medical images. By developing an automatic pipeline for the detection of spongiosis, we aim to overcome some limitations of manual feature extraction. The results demonstrate that our method correctly identifies and characterize spongiosis and allows the extraction of features that will help to better understand the spongiosis patterns in different CJD patients.

1. Introduction

Creutzfeldt–Jakob Disease (CJD) poses a challenge within the spectrum of neurodegenerative disorders. Characterized by a rapidly progressive clinical course and no known definitive cure, this condition falls within the group of transmissible spongiform encephalopathies (TSEs) or prion diseases [1]. CJD represents the most frequent form of prion disease in humans, yet it remains relatively rare, with an annual incidence of approximately one in every million individuals worldwide [2]. This disease is caused by the pathological accumulation of the prion protein (PrP) in the central nervous system [3], leading to progressive and severe deterioration of cognitive capabilities, with symptoms such as sleep disturbances, psychotic symptoms, and depression [4].

Although the incidence is relatively low [2], CJD manifests in three main categories: sporadic, hereditary, and acquired. The sporadic form, known as sCJD, is considered a spontaneous neurodegenerative disorder [2] and is the most common form, accounting for approximately 85% of the total cases of the disease. Within the sCJD category, there are six subtypes, each adding further variability and complexity to the disease [5]. The rarity of its occurrence, combined with the wide range of possible clinical presentations, makes diagnosis particularly challenging [6]. In addition to neuronal loss and gliosis, one of the main neuropathological features of CJD is spongiform change, characterized by the small round vacuoles, which may be clustered in the neuropil of the deep cortical layers, the subcortical grey matter, and the cerebellar cortex [5,7]. We must also consider that spongiosis is not exclusive to CJD [1], as it can be present in other conditions as well. While it is a symptom associated with CJD, it is not a direct or definitive indicator of the disease.

Recently, digital pathology has gained significance by enabling the curation of extensive repositories of gigapixel images of whole slide images (WSIs) [8], offering a new window into the study and diagnosis of complex diseases like CJD. In this context, our research focuses on the automatic detection and characterization of vacuoles in microscopic images of brain tissue, a key pathological indicator of CJD [5]. In this paper, we propose an approach for the identification and study of spongiosis patterns, aiming to facilitate rapid and accurate detection of spongiosis using artificial intelligence techniques for automatic interpretation of microscopy images. This method employs both image processing and graph theory.

By using a combination of digital image processing techniques and machine learning algorithms, this work presents a new method for automatically measuring the quantity of vacuoles in microscopic images. These techniques analyze the images to automatically detect and quantify vacuoles, providing a metric for the severity of spongiosis present in brain tissue samples. After identifying the vacuoles in the image, we use graph networks to extract different features that we can correlate with clinical variables. Graph analysis stands out as a groundbreaking method for examining intricate pathological patterns like those seen in spongiosis. The proposed methodology is evaluated by matching the features obtained from images with clinical variables from a cohort of 82 patients.

The organization of this paper is outlined as follows: Initially, the Section 2 provides an overview of the existing knowledge and outlines the materials and techniques essential for conducting this research. Subsequently, the Section 3 delves into the specifics of the algorithm employed for processing images and extracting and classifying features. In the Section 4, we present a selection of the data obtained through our algorithm. Finally, in the Section 5 and  Section 6, we engage in an analysis of the data and seek to draw some definitive conclusions.

2. Background and Materials

2.1. Dataset

Samples of brain tissues and data from the 64 cases with neuropathological diagnoses of CJD included in this study were provided by Biobank IIS Galicia Sur, and they were processed following standard operations with the appropriate approval of the Ethics and Scientific Committees. Additionally, Biobank provided 23 cases without neurodegenerative disease, which were used as controls.

Following established protocols, coronal sections of the left hemisphere from the cases and controls were fixed in 4% buffered formalin, were treated with formic acid, and were subsequently embedded in paraffin. Two fragments per case containing the striatum nucleus and cerebellar vermis were selected. Sections were cut at 5 $\mathsf{\mu }$m and were stained with hematoxylin and eosin. In total, 172 sections were digitized.

Images were digitized using an Aperio scanner. From each of these samples, a single .tiff file containing an image in RGB format was generated. An example of the scene is shown in Figure 1a for the cerebellar vermis and Figure 1b for the striatum nucleus. The resolution of the images measures 50,000 pixels in height and 40,000 pixels in width, with each pixel equal to 0.0625 $\mathsf{\mu }{\mathrm{m}}^2$ in actual size.

The striatum is the region that shows major spongiosis in cases of disease [9]. Figure 2 shows a piece of one of the highest-resolution images of the striatum: the image is from a case of disease, and regions of the tissue that indicate spongiosis are marked with green arrows.

As said, the spongiosis appears mainly on the striatum, and only on concrete types of regions of this tissue called the putamen and caudate nuclei [10], which in Figure 1 correspond to the lighter pink zones of the tissue, and there is little to no presence on the darker areas, which are called the internal capsule [10]. From now on, we will refer to the putamen and caudate nucleus as the tissues of interest (ToI).

For the patients with CJD, there is some information about their clinical variables such as sex, age, days of evolution or the presence of a “panencephalic form”; Supplementary Table S1 provides the whole data on these variables.

2.2. Pipeline Components for Feature Extraction

In this section, two techniques for processing data that are employed in the method are presented: namely, a Gaussian filter for image smoothing and some well-known roundness criteria. A Gaussian filter is instrumental for reducing image noise and is a useful tool to identify regions with the same color pattern, which will be useful to identify the ToI. Additionally, we employ roundness criteria to evaluate how closely the shape of our vacuole candidates approximates a perfect circle. These metrics are also vital in our study, as the structures we aim to identify are round vacuoles.

2.2.1. Gaussian Filter

A Gaussian filter operates by convolving images with a Gaussian function [11]. Mathematically, the Gaussian function in two dimensions is defined as:

$$g(x,y)={\displaystyle \frac{1}{2\pi {\sigma }^2}}{e}^{-{\displaystyle \frac{x^2+y^2}{2{\sigma }^2}}},$$

(1)

where x and y represent the distances from the origin in the horizontal and vertical dimensions, respectively, and $\sigma$ is the standard deviation of the Gaussian distribution. This standard deviation parameter ($\sigma$) controls the extent of smoothing, with larger values leading to more significant blurring effects.

By adjusting the $\sigma$ value, the filter’s sensitivity can be used with features of varying sizes and contrast levels, ensuring optimal preparation of the images for subsequent analysis stages. The Gaussian filter’s properties also ensure uniform smoothing in all directions, preserving the natural structures within the image while preparing it for the accurate identification and analysis of target regions.

2.2.2. Roundness Criteria

There are many criteria to determine the circularity of a geometric shape [12]; three of the most commonly used are:

• Circularity Ratio:

This formula quantifies the circular nature of a region, with higher values indicating shapes that more resemble a perfect circle. It is a dimensionless number that serves as a measure of roundness.

$$\mathrm{Circularity}=4\pi \times {\displaystyle \frac{\mathrm{Area}}{{\mathrm{Perimeter}}^2}}.$$

(2)

• Aspect Ratio:  

The aspect ratio evaluates the elongation of a shape, providing a measure of how much a region deviates from being equidimensional. High values suggest less elongation and indicate shapes that are more circular or square.

$$\mathrm{AspectRatio}={\displaystyle \frac{\mathrm{Major}\mathrm{Axis}\mathrm{Length}}{\mathrm{Minor}\mathrm{Axis}\mathrm{Length}}}.$$

(3)

• Fill Ratio:  

This ratio assesses the compactness of a region by comparing its area to the area of a circle with the same major axis length. It reflects the extent to which a region fills a circular area with a diameter equal to its major axis length, providing insight into the density and regularity of the shape.

$$\mathrm{FillRatio}={\displaystyle \frac{\mathrm{Area}}{\pi \times {\left(\frac{\mathrm{Major}\mathrm{Axis}\mathrm{Length}}{2}\right)}^2}}.$$

(4)

Given their definitions, it is easily demonstrable that all three functions have a range: $(0,1]$. For each function, a value approaching 1 signifies a shape that closely approximates a perfect circle.

2.3. Graph Theory for Feature Extraction

Graph theory has become a helpful tool for the representation of interconnected data and has provided significant advances in various fields, including medical imaging [13,14]. Lately, we have been witnessing the systematic organization of vast amounts of information into knowledge graphs.

Graph theory can be used for the creation, manipulation, and study of complex networks. This approach allows for the modeling of interactions and connections between different regions or points of interest in images, providing a deeper understanding of the underlying structures and patterns [13].

Graph theory offers a suite of strategies for feature extraction and enhances the understanding and interpretation of data; one of the most efficient strategies is the detection of communities in the graph [15].

Implementing algorithms such as community_louvain [16] facilitates the identification of communities or clusters within networks. In summary, communities in graphs are subsets of nodes or points that are more interconnected with each other than with the rest of the network.

This theoretical field allows us to extract some metrics that can be obtained using graphs:

• Average clustering coefficient: This metric assesses the degree to which the neighbors of a node are interconnected, offering insights into the propensity for vacuoles to cluster together [16].
• Average degree of connectivity: This metric quantifies the mean number of connections each node has within the network, offering insight into the graph’s overall connectivity density. To avoid the unrealistic scenario wherein every node is considered connected to all others, it is essential to establish a maximum distance criterion for connections. In this study, we employ the Euclidean distance as the threshold to determine the connectivity between vacuoles.
• Global efficiency: This indicator measures the graph’s overall efficiency in information transfer, illustrating how swiftly information can traverse the network from one node to another [16].
• Modularity: Modularity evaluates the graph’s division into modules or communities; a high modularity score signifies a clear delineation between densely connected clusters of nodes and their weaker inter-cluster connections [16].
• Degree assortativity coefficient (degree correlation): By examining the tendency of nodes to connect with others possessing a similar degree of connections, this measure provides insight into the structural hierarchy and clustering tendencies among the nodes [16].

3. A New Method for the Automatic Feature Extraction and Classification of Creutzfeldt–Jakob Disease

The detection of spongiosis, characterized by the presence of vacuoles in the brain tissue, is very helpful for the accurate diagnosis of CJD [5]. This paper presents a new method for the automatic identification and characterization of vacuoles in microscopic brain tissue images based on two stages. The first stage is devoted to identifying vacuoles in the images, and the second stage studies the features of these findings. These stages are detailed in Section 3.1 and Section 3.2, respectively. During the whole method, we work only with the blue components of the images, as in our samples, these prove to be the areas with the most contrast between values for the white and tissue regions.

3.1. Image Processing Pipeline

The processing pipeline is composed of two main parts: one consisting of image preprocessing (Section 3.1.1) and the second part devoted to the identification of sets of vacuoles (Section 3.1.2). A schematic view of the whole image processing pipeline is provided in Figure 3.

3.1.1. Preprocessing

Preprocessing includes a set of tasks used to obtain a “cleaned” version of the input image that is more suitable for proceeding with the analysis. In order to adapt the input to the proposed analysis scheme, we must consider that the extension of tissue in the image is variable between images, and some tissue samples may display minor tears made during the extraction of the tissue (see Figure 1b for an example); furthermore, the sizes of the ToI vary depending on the image. Due to this fact, a step to correct this situation by eliminating the darker areas and the white regions in order to work only with the ToI is required. We propose applying a Gaussian filter (see Section 2.2.1), with $\sigma =1$ in Equation (1). We continue by segmenting the filtered image into three zones based on intensity: dark areas (normalized intensity below 0.25), white areas (normalized intensity above 0.9), and neutral zones (intensity between 0.25 and 0.9). We then focus on the neutral region (intensity between 0.25 and 0.9), which we consider as the ToI. This step provides a region of interest, ensuring the method only focuses on areas prone to exhibiting spongiosis.

Due to the Gaussian filter, small objects (artifacts) could appear in both dark and white areas. This leads to errors when artifacts in the white or dark areas are mistakenly identified as potential vacuoles; such small objects must be removed from the white and dark areas. We maintain the small white and dark regions as part of the ToI and only intend to exclude the large white or dark regions.

The output of this step is shown in Figure 4, which is a segment from the ToI.

This completes the identification of the neutral zone by selecting areas that are neither dark nor white, effectively capturing regions of intermediate intensity. After this process, we work with the neutral zone as our effective area.

Figure 5 shows the final result of the preprocessing procedure applied to one of the images.

3.1.2. Identification of a Set of Vacuoles

Once we obtain the zones where spongiosis can be present, we can search for structures similar to the vacuoles, which are round holes as explained in Section 1. As vacuoles are holes with no tissue [5], they appear as white microregions in the blue component of the images employed in the pipeline. As the purpose is to identify such holes, we set specific intensity thresholds to distinguish between dark and white regions of the image.

To detect vacuole candidates, we can start by searching the white regions using the following sequence of tasks: first, we calculate the average value of the pixels in the four corners of the image, as all images have no tissue in these areas. In this work, we take a square of $1000\times 1000$ pixels for each corner. After obtaining the mean value of these pixels, which is referred to as the CornerMean, the following threshold can be used:

$$Threshold:=CornerMean·0.3+ImageMean·0.7,$$

(5)

where $ImageMean$ is the mean value for the pixels of the blue component of the whole image. With this number, we try to ensure that we obtain a threshold value between the value of the ToI and the value of the white regions.

The white regions, which are potentially indicative of vacuoles, and the colored regions, which correspond to the surrounding tissue, are separated by labels for further analysis. Following the determination of the threshold values, the filtered image undergoes segmentation and labeling. This process identifies sets of connected pixels within the same labeled object. Segmentation, which is used in image processing to separate different regions based on pixel intensity values, utilizes the previously determined intensity thresholds to separate and classify the image pixels that fall above or below these thresholds into two separate groups: vacuole candidates and non-vacuole candidates. During labeling, connected pixels from vacuole candidates are merged into objects. Small objects that do not meet a minimum size (ms) are removed from the vacuole candidates, improving the specificity of the analysis by focusing on relevant structures and discarding artifacts.

In this work, for the minimum size (ms), all objects must have 45 or more pixels (2.8125 $\mathsf{\mu }$${\mathrm{m}}^2$) to be considered candidates for vacuoles. We chose this value as in our trials, every vacuole candidate with lesser size was noise or not a real vacuole.

In our methodology, vacuole identification is refined through morphometric analysis. Applying the circularity criteria explained in Section 2.2.2, we impose the condition that all the candidates for vacuoles must have a minimum value in all three Equations (2)–(4). All vacuole candidates that do not exceed a score of 0.7 for the fill, aspect, and circularity ratios (defined in Section 2.2.2) are discarded. The output of this process is the set of vacuoles in Figure 6.

3.2. Feature Extraction

Once the vacuoles are obtained from the processed images, as described in Section 3.1, a set of features that allow us to characterize the identified vacuoles can be extracted. These features include two different approaches that use geometrical and graph features.

Graph features have already been explained in detail in Section 2.3. In contrast to graph features, geometrical features provide an analysis of the set of vacuoles, focusing on their quantity and dimensions rather than their behavior as a group. The set of proposed geometrical features utilized in this work comprises two values:

• Normalized area: the area of the vacuoles relative to the total tissue area in the image, providing an indicator of the severity of the spongiosis.
• Total number of vacuoles: the number of vacuoles found in the image.

Graph analysis emerges as an innovative technique for the study of complex pathological patterns such as spongiosis in CJD. By modeling the interactions and spatial relationships between the identified vacuoles in brain tissue, graphs allow us to gain a deeper understanding of the organization and distribution of these pathological structures at a microscopic level. To obtain a graph-type structure from our image, first, we need to define which elements will be considered nodes.

• Creation of Vacuole-Based Graphs:  

Each identified vacuole is represented as a node in the graph. Edges between nodes are established based on spatial proximity and shared morphological features, allowing modeling of the connectivity and clustering of vacuoles within the tissue. Once the graph is created, graph features can be obtained.

•   Graph Feature Extraction and Topology Analysis:

• Community detection: Through community detection algorithms, we identify groups of vacuoles that show significant clustering patterns. This clustering may indicate areas of greater affection by or severity of spongiosis, providing valuable clues about the progression of CJD.
• Centrality and graph density: We analyze centrality measures to identify nodes (vacuoles) that play critical roles in the structure of the affected tissue. The graph’s density offers information about the generality of spongiosis in the analyzed sample.

The main features obtained through graphs are detailed in Section 2.3, and we use the centroids of our vacuoles as the nodes in our graph. It is important to note that for the average degree of connectivity feature, we consider two nodes connected if their Euclidean distance is less than 1000, with pixels being our unit of measurement. It was determined that a value of 1000 gives a good balance between the computational cost and the information obtained using the algorithm. Figure 7 presents a detailed scheme of the feature extraction process.

3.3. Coding Details

All codes are implemented in Python 3.6 [17]. For Gaussian smoothing, we employ filters.gaussian(imageB) from the scikit-image library [18] to reduce noise, where imageB represents the blue component of the original image. After applying the Gaussian filter, the algorithm segments the image into three zones based on intensity. For object removal, we specify a minimum value of 0.5% of the total pixels to retain significant features, and we use remove_small_objects from skimage.morphology [18].

For the minimum size (ms) during artifact removal, all objects must have 45 or more pixels to be considered candidates for vacuoles.

Finally, for vacuole identification, all vacuole candidates that do not exceed a score of 0.7 for the fill, aspect, and circularity ratios (defined in Section 2.2.2) are discarded.

4. Results

This section presents the results of the entire process. These results evaluate the validity of the proposed method by correlating the features extracted from the images with the clinical variables of the patients.

4.1. Spongiosis Segmentation

The first experiment was conducted with the aim of evaluating the accuracy of the proposed methodology in terms of comparing case and control patient images. The pipeline was applied to the whole set of images, and the results were grouped into two separate groups. The results from this algorithm show that spongiosis is less present in cases of the control than in the CJD group: the spongiosis percentage measure (the geometrical feature: normalized area) has a mean of 0.0132 for the control group and 0.1748 for the CJD cases group. Figure 8 provides a graphical plot of the box and whiskers diagrams of the normalized areas for both the CJD cases and the control patients:

4.2. Feature Extraction

This section presents the results of the second phase of our work, which is devoted to extracting features from the previously identified vacuoles. The features extracted are the ones mentioned in Section 2.3 and Section 3.2.

Table 1. Table of the means and standard deviations of the geometrical features of both groups: the control group and the CJD cases.

Group	Normalized Area Mean	Normalized Area Std	Total Number of Vacuoles Mean	Total Number of Vacuoles Std
Control group	0.0128	0.0120	1868.08	1781.64
CJD cases	0.1778	0.1548	14,079.36	12,459.91

sums up the geometrical feature results (the complete results are given in Supplementary Table S1).

Now

Table 2. Table of the means of the graph theory features for both groups: the control group and the CJD cases.

Group	Average Clustering Coefficient	Average Degree of Connectivity	Global Efficiency	Modularity	Degree of Correlation
Control group	0.5129	9.9677	0.0531	0.7532	0.7319
CJD cases	0.6091	85.6109	0.0792	0.8485	0.8433

shows the means of the graph theory features presented in Section 2.3 is given (the complete results are given in Supplementary Table S1).

4.3. Correlation of Features with Clinical Variables

After obtaining the features, it is now possible to search for correlations between them and the clinical variables involved in the study. The following clinical variables are considered: age, sex, the presence of “panencephalic form”, and days of CJD evolution, where “days of CJD evolution” refers to the number of days elapsed since the onset of the first symptoms.

These correlations are only made with cases of the disease and not with patients from the control group. The first two clinical features to be analyzed are age and days of evolution, both of which are numerical. In contrast, sex and the presence of “panencephalic form” are categorical features. Due to these differences, the procedures for analyzing each type of feature varies accordingly.

4.3.1. Age and Days of Evolution

In order to analyze the results and their correlations with the clinical numerical variables (age and days of evolution), the Pearson correlation [19] was employed, as it is one of the most widely used technique for evaluating the correlation between two quantitative variables in biomedical research. The Pearson correlation provides r values ranging from −1 to 1, with −1 indicating inverse correlation and 1 indicating direct correlation. The results show that there is a strong inverse correlation between some features and the days of evolution; there is also correlation between age and some of the features. Figure 9 provides the heatmap of the Pearson correlations between the two clinical variables and the different metrics obtained with the previously explained algorithm:

4.3.2. Sex and Panencephalic Form

In order to analyze the results and their correlations with the clinical categorical variables (sex and presence of panencephalic form), a radar chart is used. The radar charts are shown in Figure 10, the means are normalized by the higher value of the compared categories:

5. Discussion

CJD is a devastating neurodegenerative disorder that requires all efforts in research disciplines to enhance our understanding and develop effective tools to improve our knowledge of the disease. The arrival of digital pathology has provided new insights and has gained significance by enabling the curation of extensive repositories of gigapixel images of WSIs. This offers a new window for studying and diagnosing complex diseases like CJD. However, the high resolution of these images results in very large file sizes, posing challenges due to the computationally costly methods required for processing.

In this paper, we propose an approach for the identification and classification of spongiosis patterns, aiming to facilitate rapid and accurate assessment of vacuoles by using artificial intelligence techniques for automatic interpretation of microscopy images. The analysis of brain tissue images through digital processing and graph theory in the study of CJD has revealed significant insights into the disease’s pathology. When it comes to identifying CJD, our method shows significant differences in the results between control and CJD patients.

By using a combination of digital image processing techniques and machine learning algorithms, this work proposes a new method for automatically measuring the quantity of vacuoles in microscopic images. These techniques analyze the images to automatically detect and quantify vacuoles, providing a metric for the severity of spongiosis present in brain tissue samples. After identifying the vacuoles in the image, we proceed to use graph networks to extract different features that we can correlate with clinical variables. Graph analysis stands out as a groundbreaking method for examining intricate pathological patterns like those seen in spongiosis.

This methodological approach was employed to analyze a cohort of 82 patients. Our findings also underscore an inverse correlation between specific features from our algorithm and the progression of CJD. Notably, some sex differences were observed, although they are not significant: it appears that spongiosis is more present in male patients. It is also clear that our study shows the lesser presence of spongiosis in the non-panencephalic forms, which is an interesting relationship.

All the parameters were chosen after multiple trials, and we considered the selected values to be the most suitable options for accurately identifying the ToI and the vacuoles. For the ToI, multiple parameters were involved. For the Gaussian filter, we chose $\sigma =1$ by default, although other values could have been selected, with the other parameters altered accordingly. For the segmentation into three zones, we chose 0.25 and 0.9 as thresholds for the dark and white areas, respectively. We selected a value close to 1 for the white regions since those zones are nearly white. Significantly increasing or decreasing these values would cause the algorithm to incorrectly detect some regions. Finally, we specified a minimum size of 0.5% of the total pixel area to retain significant features. This value can be slightly adjusted, and the results would be similar. Its importance lies in ensuring that vacuoles are not mistakenly labeled as part of the white region, as we are trying to locate them in the ToI.

In all images, we found small dots or micro-scars in the brain tissue that were clearly not vacuoles; therefore, we chose a minimum size of 45 pixels (2.8125 $\mathsf{\mu }$${\mathrm{m}}^2$) during artifact removal. Increasing the size would be problematic since some vacuoles are very small, with sizes of 70 pixels (4.375 $\mathsf{\mu }$${\mathrm{m}}^2$) or less; decreasing the size drastically would make the algorithm detect some of the dots or micro-scars as vacuoles.

During vacuole identification, we chose 0.7 for all three roundness parameters. In our trials, these values proved effective. Lowering them resulted in the algorithm detecting many non-vacuole objects, while increasing them caused the algorithm to discard real vacuoles.

The application of image processing techniques and graph networks in this study demonstrates the potential for such methodologies to uncover patterns in neurodegenerative diseases. Limitations of our study, particularly the sample size, highlight areas for further investigation. Future research could focus on expanding the dataset and exploring additional neurodegenerative diseases.

The implications of our findings extend beyond CJD to offer a perspective on the study of neurodegenerative diseases. By leveraging the power of digital image processing and graph theory, we open the door to more analyses of disease markers, which could lead to better understanding of these conditions. Furthermore, the method’s adaptability suggests it could be applied to a wide range of neurodegenerative disorders, offering a versatile tool for research and clinical diagnosis.

As mentioned in the introduction, sCJD has six subtypes. Due to the limited size of our cohort—resulting from the rarity of CJD—we opted to analyze the disease as a whole, as in most instances, we lacked a statistically significant number of cases within these smaller subgroups. We encourage future researchers to explore the potential outcomes of applying these techniques to the different subgroups of sCJD. We also acknowledge that our control group does not represent the real-world challenges of CJD diagnosis, as it could benefit from being larger and including a broader range of conditions. However, given that our primary objective was to detect and study spongiosis, we believe this dataset is adequate for that purpose.

Further development of the methodology in order to enhance its sensitivity and specificity as well as expanding the dataset—by increasing the variability and size within either the CJD group or the control group—would be valuable steps towards maximizing the impact of this research. This algorithm is also highly adaptable and useful for a wide range of applications. Its primary objective is to identify round, colored structures based on specific criteria. With a few modifications, this algorithm can be modified to detect other similar conditions, either in human or animal tissue samples. The ultimate goal is to establish a robust and applicable tool that can significantly contribute to the neurodegenerative disease field, offering insights that could lead to breakthroughs in the understanding of these complex conditions.

6. Conclusions

This work introduces a novel method combining artificial intelligence and graph theory to analyze histopathological images of CJD. By using image processing and machine learning techniques, the research focuses on the automated detection and quantification of spongiosis. This approach addresses the limitations of manual feature extraction, facilitating better understanding of spongiosis patterns in disease patients. The methodology demonstrates effective identification of spongiosis, offering insights into the disease’s pathological features, and is a promising approach. This method could be extended to other neurodegenerative image characterization tasks as well. Our research demonstrates the utility of this methodology for automatizing spongiosis detection and the understanding of CJD through the quantitative analysis of spongiosis in the brain tissues. The significant correlations found between the features extracted from the images and the clinical variables underscores the potential of our approach to contribute valuable insights into the progression and severity of CJD.