The Role of Artificial Intelligence and Neural Networks in Developing Digital Healthcare Applications

Prepared by the researche  : Dr. Suzan Saber Haydar¹ , Khalan Jalil Rostam²

• ^{1, 2} Sulaimani University, College of Administration and Economics

DAC Democratic Arabic Center GmbH

International Journal of Economic Studies : Thirty-fifth Issue – November 2025

A Periodical International Journal published by the “Democratic Arab Center” Germany – Berlin

:To download the pdf version of the research papers, please visit the following link

https://democraticac.de/wp-content/uploads/2025/11/%D8%A7%D9%84%D9%85%D8%AC%D9%84%D8%A9-%D8%A7%D9%84%D8%AF%D9%88%D9%84%D9%8A%D8%A9-%D9%84%D9%84%D8%AF%D8%B1%D8%A7%D8%B3%D8%A7%D8%AA-%D8%A7%D9%84%D8%A7%D9%82%D8%AA%D8%B5%D8%A7%D8%AF%D9%8A%D8%A9-%D8%A7%D9%84%D8%B9%D8%AF%D8%AF-%D8%A7%D9%84%D8%AE%D8%A7%D9%85%D8%B3-%D9%88%D8%A7%D9%84%D8%AB%D9%84%D8%A7%D8%AB%D9%88%D9%86-%D8%AA%D8%B4%D8%B1%D9%8A%D9%86-%D8%A7%D9%84%D8%AB%D8%A7%D9%86%D9%8A-%E2%80%93-%D9%86%D9%88%D9%81%D9%85%D8%A8%D8%B1-2025.pdf

Introduction

Carcinogenesis refers to uncontrolled cell development caused by the silencing of tumour suppressor genes or the activation of oncogenes. Epigenetic changes, together with genetic modifications, influence the development of cancer. Cancer is defined by uncontrolled cell growth that can manifest in various degrees of severity, ranging from mild to life-threatening [1]. Developmental problems can arise from mutations in crucial cellular genes, which serve as a primary conduit for the transformation of proto-oncogenes into their oncogenic form. Cancer formation is significantly influenced by the gradual accumulation of many mutations throughout an individual’s lifespan [1, 2]. Epigenetic modifications are heritable differences in gene expression that do not result from changes in the DNA sequence. Examples of these modifications include DNA methylation and chromatin regulation [3]. CRC ranks fourth in terms of frequency and third in terms of cancer-related deaths globally [4, 5]; CRC incidence and mortality rates have increased dramatically worldwide. It is expected that by 2035, the number of new cases and fatalities from colon cancer and rectal cancer will increase by 60% and 71.5%, respectively [6, 7]. Genetic mutations affecting the epithelial tissues of the colon and rectum lead to CRC, with a particular emphasis on genes associated with DNA repair mechanisms, tumour suppressors, and oncogenes. Genetic and epigenetic instability are significant factors in the onset and progression of CRC. Colorectal cancer pathogenic mechanisms may involve three distinct molecular pathways: microsatellite instability (MSI), chromosomal instability (CIN), and the cytosine preceding guanine (CpG) island methylator phenotype (CIMP) pathways [8]. Deep learning has emerged as a revolutionary application in medical imaging, moving beyond traditional algorithms to systems that learn directly from data. In this context, “learning” is the process of converting vast amounts of experiential input—thousands of labelled medical scans—into operational knowledge or diagnostic competence. This means the algorithm progressively improves its ability to identify patterns, such as tumours or anomalies, without being explicitly programmed to look for specific shapes or densities, fundamentally changing how we extract information from images. The cornerstone of this advancement is the CNN, a specialized architecture whose design is inspired by the hierarchical and localized receptive fields of the primary visual cortex in the brain. This biological analogy is crucial; just as our visual system processes edges and shapes before assembling them into complex objects, CNNs build understanding through layers of abstraction. Their superiority became undeniable when they consistently began to outperform conventional computer vision techniques, marking a pivotal shift in the capabilities of automated image analysis. The network’s power derives from its use of the convolution operation to hierarchically decode intricate patterns within a set of images. This operation is built upon two fundamental components: the input image (a grid of pixel values) and a kernel or filter (a small matrix of weights). The filter systematically slides (or “convolves”) across every possible position in the input image. At each location, it performs an element-wise multiplication between the filter’s weights and the underlying pixel values of the current image segment. The products of these multiplications are then summed into a single, weighted value. This value is placed in a corresponding position in a new output array, ensuring the spatial structure of the features is preserved. This sliding and calculating procedure is repeated across the entire image, generating a new processed output called a feature map. Each feature map’s values are then adjusted by adding a bias term, which allows the model to shift the output. Subsequently, each element in the map is passed through a non-linear activation function (like ReLU). This critical step introduces non-linearity into the system, enabling the network to learn and represent complex, non-linear relationships in the data that a simple linear operation could not, such as distinguishing between highly irregular textures. In a CNN, this entire process is repeated across multiple layers. The weights within the filters are not pre-defined but are the key trainable parameters the network learns during training. Through a process called backpropagation, the network discovers the optimal filter values that best highlight the features necessary for the intended task (e.g., identifying a specific disease). Early layers learn to detect simple, low-level features like edges and blobs. The outputs of these layers (feature maps) are then fed as inputs to subsequent layers, which combine these simple features to detect more complex, high-level patterns and structures. This repetition builds a deep hierarchy of abstraction, from simple edges to entire anatomical structures, allowing for sophisticated interpretation [9,10]. In this paper, we aimed to train the DenseNet-201 algorithm, one of the CNN architectures, and a technique of DL, which is a part of AI, to classify strips of CRC patients’ cancer on the DNA strip. The science and engineering of creating clever devices and brilliant computer programs is known as AI. It connects to the related problem of employing computers to analyse human intelligence. The methods address relatively straightforward or intricate issues that arise within more intricate systems. While some works in AI simulate human intelligence, most focus on analysing the challenges intelligence faces.

1– Machine Learning

Machine learning provides the essential framework that allows CNNs to automatically learn and extract meaningful features from visual data. Unlike traditional image processing methods that rely on manually designed filters and algorithms, CNNs utilize supervised learning to iteratively adjust their internal parameters—primarily the weights of their convolutional kernels—through processes like backpropagation and gradient descent. This enables the network to discern hierarchical patterns, starting from low-level features such as edges and textures in earlier layers, to increasingly complex and abstract structures like shapes and objects in deeper layers. By training on large labelled datasets, the CNN optimizes these filters to minimize prediction error, effectively learning the most discriminative features for tasks such as image classification, object detection, and segmentation. Thus, machine learning transforms the CNN from a fixed architecture into an adaptive model capable of generalizing from examples and performing sophisticated visual recognition tasks with high accuracy [11,12].

1.2 Deep Learning

Deep Learning serves as the foundational engine that empowers CNNs to autonomously learn and represent complex hierarchical features from visual data. Unlike traditional machine learning methods that often depend on manually curated features, deep learning enables CNNs to process raw pixel inputs and progressively abstract them through multiple layers of non-linear transformations. In a typical CNN architecture, initial layers capture low-level features such as edges and textures, while deeper layers synthesize these into high-level constructs like object parts and entire categories. This multi-layered hierarchy is optimized end-to-end via backpropagation and gradient-based learning, allowing the network to refine thousands of parameters—primarily convolutional filters—to minimize prediction error. By leveraging large-scale datasets, deep learning equips CNNs with the capacity to generalize across diverse visual tasks, from image classification to semantic segmentation, establishing them as a cornerstone of modern computer intelligence [13,14].

1.3 Feed Forward Neural Networks

An artificial neural network with a feed-forward topology is called a Feed-Forward Neural Network (FFNN) and, as such, has only one condition: information must flow from input to output in only one direction with no back-loops. There are no limitations on the number of layers, the type of transfer function used in individual artificial neurons, or the number of connections between individual artificial neurons. Within the deep learning architecture of a Convolutional Neural Network (CNN), feedforward neural networks, specifically fully connected (FC) layers, serve as the final stage for high-level reasoning and classification. The initial convolutional and pooling layers act as automated feature extractors, transforming the raw input image into a rich set of feature maps that encode hierarchical patterns. However, these features are still spatially structured. The FC layers accept these flattened feature vectors and perform a traditional feedforward operation, where every neuron is connected to all activations from the previous layer. Through a series of weighted sums, bias additions, and non-linear activation functions, these layers learn to combine the distributed features into a global representation and map them to the final output classes, such as image categories. Thus, while convolution handles the spatial feature learning, the feedforward components are crucial for synthesizing those features into a definitive prediction, showcasing the hybrid and complementary nature of these architectures within deep learning [15].

1.4 Categorical Cross-Entropy (CCE)

In CNNs designed for multi-class classification, Categorical Cross-Entropy serves as the fundamental loss function that quantifies the disparity between the predicted class probability distribution and the true, one-hot encoded label. It calculates the negative sum of the true labels multiplied by the logarithm of the corresponding predicted probabilities, effectively measuring how well the model’s SoftMax-activated output, which represents a probability score for each class, aligns with the ground truth. By penalizing confident but incorrect predictions more heavily than hesitant wrong ones, this loss function provides a strong, differentiable signal that guides the optimization process during backpropagation. Minimizing Categorical Cross-Entropy is therefore essential for training a CNN to not only make accurate classifications but also to output well-calibrated probability distributions that reflect its certainty [16]. The CCE loss function demands maximizing the log-likelihood of the N-sample training set, while yi is the index of the actual class for the sample i as the following

Where :  is a d-dimensional vector,  is the probability of  Matching  is the learned parameters of the classifier.

1.5 Back propagation algorithm

The backpropagation algorithm in a Convolutional Neural Network is the essential iterative process that enables learning by calculating the gradient of the loss function with respect to every weight in the network. It begins with a feedforward pass, where an input image is progressively transformed through convolutional layers that extract features using learnable filters, activation functions like ReLU that introduce non-linearity, and pooling layers that reduce spatial dimensions, ultimately culminating in a set of class probability predictions from a fully-connected output layer. Following this, the loss, often Categorical Cross-Entropy, is computed to measure the error between the prediction and the true label. The algorithm then performs the backward pass, recursively applying the chain rule from the final layer back to the first to distribute the calculated error gradient backwards through each operation: through the fully-connected layers, up sampling through the pooling layers, and, most critically, back to the individual weights of the convolutional filters themselves. This calculated gradient is then used by an optimizer like SGD or Adam to update the filter weights and other parameters, finely tuning them to minimize the loss and improve the network’s feature extraction and classification accuracy on subsequent iterations [11,15]. The backpropagation equation for deltas is

1.6 Convolutional Neural Network

A Convolutional Neural Network (CNN) is a deep learning architecture specifically designed for processing structured, grid-like data such as images, leveraging a sequence of specialized layers to automatically and adaptively learn spatial hierarchies of features. The core of this architecture is the convolutional layer, which employs learnable filters that perform convolution operations by sliding across the input data. These filters detect local patterns—such as edges, textures, and shapes—with the step size, or stride, determining the spatial movement and influencing the output size and receptive field overlap. Subsequent pooling layers (e.g., max pooling) down sample the feature maps, reducing spatial dimensions, enhancing translational invariance, and decreasing computational load. After repeated convolutional and pooling operations, the resulting high-level features are flattened and passed to fully connected layers, which integrate these distributed features to perform classification or regression. The evolution of CNN architectures has introduced increasingly deep and efficient designs, such as AlexNet, VGGNet, and ResNet, each addressing challenges like training stability and feature propagation. A notable advancement is DenseNet (Densely Connected Convolutional Network), which connects each layer to every other layer in a feed-forward manner within dense blocks. This design promotes feature reuse, alleviates the vanishing gradient problem, and improves parameter efficiency by encouraging the flow of information and gradients throughout the network [12,14]. In the applied side of this paper, the DenseNet201 was employed to classify DNA imaging for CRC patients.

2- Method and tools

The data sample consisted of 537 patients and included 27,525 images cropped from the DNA CRC strip. The data were collected from Hiwa Hospital, Sulaimani, from November 1, 2021, to January 1, 2025. The image is classified into eight classes (BAT25, BAT26, BAT40, NR21, NR22, NR27, D2S123, and D5S346) depending on microsatellite CRC type, diagnosed from the DNA strip. From here, the image in the first class consists of 8347 images, and the image is classified into seven classes (NR27, D2S123, D5S346, D17S250, Mycl1, TPOX, and TH01) depending on the microsatellite CRC types. The second class consists of 19366 images, which are classified into eight classes depending on the microsatellite CRC types. These eight/seven classes will be split into training, validation, and testing. The training set takes 70% of the images, the validation set takes 20%, and the testing set takes only 10% of the images.  The input data will be split into training, validation, and testing. The images must be the same size when inputting them into DenseNet-201. All images are input as 2D. Also, a batch size equal to 64 was used. An Early stop has been used for the training set to reduce overfitting, and the best weights are saved for the training iteration. The model of DenseNet-201 was called from the Keras library and set to classify eight classes and a few cases into seven classes. After that, the model starts to train and continues for 40 epochs, and the Adam Optimizer has been used with a learning rate = 0.0001. In addition to using a categorical loss function. The time duration of training is from four hours to 24H. From training each training set and validation set, acceptable accuracy is obtained for each of the fourteen classes, and can classify eight/seven types of CRC.

Fig.1 : Shows the model of the Dense Net.

3- Results and their discussion

After training the network for 40 epochs, the result of training is shown in Table 1.

Table 1: The DenseNet-201 Accuracy

Based on the results in Table 1, the DenseNet-201 model completed 40 epochs and achieved a Training Accuracy of 98.12%. The network successfully categorized images into eight distinct classes, with a high performance of 98.09% across these groups. The Validation set yielded scores of 90.32% and 90.11%, with the highest performance observed in the normal BP group at 90%. Conversely, the Hypertension class showed comparatively lower results within the Validation set. For the Testing set, outcomes were 90% and 89%, with normal BP again demonstrating the strongest result at 90.20%, while Hypertension recorded 89.55%. The Hypertension class consistently showed lower performance across the Training, Validation, and Testing sets. This pattern suggests that increased sample size may correlate with a reduction in performance, necessitating extended model development time. In summary, the evaluation of DenseNet-201 indicates acceptable overall performance for the data used.

4- Conclusion

Colorectal cancer, referred to as cancer of the large intestine, is broadly defined in this context to encompass cancers occurring in the colon, the rectosigmoid junction, and the rectum. While there may be exceptions, cancers of the colon typically represent the majority, constituting around two-thirds of the total cases, while cancers of the rectosigmoid junction and the rectum make up the remaining one-third. Artificial Intelligence (AI) and Artificial Neural Networks (ANNs) are transforming digital healthcare by improving diagnostic accuracy, optimizing treatment strategies, and facilitating real-time patient monitoring. Colorectal cancer (CRC), a complex and molecularly heterogeneous disease, presents varied clinical behaviours influenced by its diverse characteristics. We provide a comprehensive review of the integration of AI and ANNs, with a specific emphasis on Convolutional Neural Network (CNN) models within modern healthcare systems. In this study, a DenseNet architecture will be trained using DNA strip samples obtained from CRC patients. The dataset comprises 27,525 cropped images derived from samples of 537 patients, collected at Hiwa Hospital in Sulaimani between November 1, 2021, and January 1, 2025. From an applied perspective, we demonstrate that the CNN architecture proved to be an effective approach for classifying the seven or eight distinct categories of colorectal cancer (CRC) types. Finally, we found that the DenseNet has high accuracy for both classes’ training and validation sets and acceptable accuracy for the testing set, and the CNN architecture was a good approach for classifying seven/eight classes of CRC cancer types. We are recommended to evaluate other state-of-the-art architectures, such as Vision Transformers (ViTs), Capsule Networks, or hybrid CNN-Transformer models, for CRC classification. These may capture more complex feature interactions and improve classification performance, especially for highly heterogeneous cancer types.

Reference

1. Kontomanolis, E. N., Koutras, A., Syllaios, A., Schizas, D., Mastoraki, A., Garmpis, N., … & Fasoulakis, Z. (2020). Role of oncogenes and tumor-suppressor genes in carcinogenesis: a review. Anticancer research, 40(11), 6009-6015.
2. Hahn, W. C., & Weinberg, R. A. (2002). Rules for making human tumor cells. New England Journal of Medicine, 347(20), 1593-1603.
3. Sadida, H.Q., et al., Epigenetic modifications: Key players in cancer heterogeneity and drug resistance. Transl Oncol, 2024. 39: p. 101821.
4. Hossain, M. S., Karuniawati, H., Jairoun, A. A., Urbi, Z., Ooi, D. J., John, A., … & Hadi, M. A. (2022). Colorectal cancer: a review of carcinogenesis, global epidemiology, current challenges, risk factors, preventive and treatment strategies. Cancers, 14(7), 1732.
5. Xi, Y., & Xu, P. (2021). Global colorectal cancer burden in 2020 and projections to 2040. Translational oncology, 14(10), 101174.
6. Araghi, M., Soerjomataram, I., Jenkins, M., Brierley, J., Morris, E., Bray, F., & Arnold, M. (2019). Global trends in colorectal cancer mortality: projections to the year 2035. International journal of cancer, 144(12), 2992-3000.
7. Tiruneh, Y. M., Beshah, D. T., & Wassie, M. (2022). Incidence of Mortality and Associated factors among Colorectal Cancer patients at Oncology Units of Northwest Ethiopia: a retrospective cohort study. Cancer Management and Research, 1445-1455.
8. Al-Joufi, F. A., Setia, A., Salem-Bekhit, M. M., Sahu, R. K., Alqahtani, F. Y., Widyowati, R., & Aleanizy, F. S. (2022). Molecular pathogenesis of colorectal cancer with an emphasis on recent advances in biomarkers, as well as nanotechnology-based diagnostic and therapeutic approaches. Nanomaterials, 12(1), 169.
9. Glover, B. M., Buckley, O., Ho, S. Y., Sanchez-Quintana, D., & Brugada, P. (2016). Cardiac anatomy and electrophysiology. In Clinical handbook of cardiac electrophysiology (pp. 1-37). Cham: Springer International Publishing.
10. LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (2002). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324.
11. Ghatak, A. (2019). Deep learning with R. Singapore: Springer Singapore.
12. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90.
13. Kim, P. (2017). Matlab deep learning. With machine learning, neural networks and artificial intelligence, 130(21), 151.
14. Elzain, H. E., Chung, S. Y., Senapathi, V., Sekar, S., Park, N., & Mahmoud, A. A. (2021). Modeling of aquifer vulnerability index using deep learning neural networks coupling with optimization algorithms. Environmental Science and Pollution Research, 28(40), 57030-57045.
15. Rostam, K. J., & Haydar, S. S. (2023). Optimal Production Decision Making by using Artificial Neural Networks and Fuzzy Linear Programming. Al-Iraqia University Refereed Journal, 57.
16. Gana, K., & Broc, G. (2019). Structural equation modeling with lavaan. John Wiley & Sons.