Colorectal Cancer Detection Tool Developed with Neural Networks

Abstract

In the last two decades, there has been a considerable surge in the development of artificial intelligence. Imaging is most frequently employed for the diagnostic evaluation of patients, as it is regarded as one of the most precise methods for identifying the presence of a disease. However, a study indicates that approximately 800,000 individuals in the USA die or incur permanent disability because of misdiagnosis. The present study is based on the use of computer-aided diagnosis of colorectal cancer. The objective of this study is to develop a practical, low-cost, AI-based decision-support tool that integrates clinical test data (blood/stool) and, if needed, colonoscopy images to help reduce misdiagnosis and improve early detection of colorectal cancer for clinicians. Convolutional neural networks (CNNs) and artificial neural networks (ANNs) are utilized in conjunction with a graphical user interface (GUI), which caters to individuals lacking programming expertise. The performance of the artificial neural network (ANN) is measured using the mean squared error (MSE) metric, and the obtained performance is 7.38. For CNN, two distinct cases are under consideration: one with two outputs and one with three outputs. The precision of the models is 97.2% for RGB and 96.7% for grayscale, respectively, in the first instance, and 83% for RGB and 82% for grayscale in the second instance. However, using a pretrained network yielded superior performance with 99.5% for 2-output models and 93% for 3-output models. The GUI is composed of two panels, with the best ANN model and the best CNN model being utilized in each. The primary function of the tool is to assist medical personnel in reducing the time required to make decisions and the probability of misdiagnosis.

1. Introduction

Over the course of a decade, an increasing number of studies on cancer detection have been published, a development largely driven by the high mortality rate of the disease [1]. The present study is focused on colorectal cancer. According to the most recent statistics, this type of cancer has the second-highest mortality rate, with lung cancer ranking at the top [2]. The danger of CRC lies in the ambiguous symptoms, which can easily be mistaken for other less severe diseases [3], and the possibility to misdiagnose the patient. Although early diagnosis remains the most important determinant of survival in CRC, limitations and obstacles have been reported. Research has demonstrated that colonoscopies are associated with a reduced probability of survival. This outcome is attributable, at least in part, to the necessity for image interpretation by the examining physician, which can lead to an overreliance on the right colon location during the diagnostic procedure [4]. In addition, a meta-analysis by Van Rijn et al. recently reported a 21% prevalence of polyps [5]. Early and precise detection are therefore required to assist the diagnosis of CRC patients and increase their survival [6].

In the review [7], different approaches for cancer detection are studied. In the context of CRC, a wide array of CNN-based methodologies exists, incorporating or not incorporating additional techniques to enhance performance. In the case of the studies [8,9,10,11,12], simpler CNN structures are trained with variable accuracy, ranging between 60% and 99.86%. Studies [10,13,14,15,16,17] utilized either special architectures or techniques like the wavelet transform or AutoML to enhance the performance of their models. In these cases, the performance was reported to elicit a wider interval, with accuracy ranging from 49% to 96%. Lastly, some researchers use ANN with different types of inputs [18,19,20,21,22]. In these cases, the performances go from as low as 63% to as high as 93%. This finding demonstrates that a satisfactory model can be trained even with a limited set of data, underscoring the potential for effective machine learning with minimal input. In the aforementioned models, either images are utilized as inputs or biological parameters are derived from blood tests or stool analysis. However, a comprehensive diagnosis necessitates the incorporation of both biologic test results and, when indicated, a colonoscopy. Moreover, the analysis of a broader array of diseases is contingent upon the identification of shared symptoms and established detection methodologies. In the same review, different types of carcinoma and diagnosis tools are studied.

For lung cancer, CNN or other neural network-based models are used in conjunction with a performance-enhancing technique. For example, refs. [23,24] use PNN and have performances of 92−95%. In the study [25], CNN is used with 97% accuracy. In research [26,27,28,29], ANN-based models are trained. The best results come from [26], where an accuracy of 97.86% is obtained. In this scenario, AdaBoost is employed in conjunction with an ANN and the enose concept is explored. In [30], DITNN is utilized, and an accuracy of 98.4% is obtained.

For pancreatic cancer, mostly ANN-based methods are studied. In [29,31,32], the performances are between 71% and 85%. The most optimal models are obtained with MNN, a combination of CNN and ANN for the same purpose. The precision of the model is determined to be 95% during the validation process.

For the diagnosis of thyroid cancer, several CNN-type methods exhibit a range of performance rates, with accuracies reported between 83% and 99%. The most optimal outcomes are achieved through the utilization of GoogleNet-based CNNs. A substantial corpus of research has demonstrated the efficacy of ANN-based models, with reported accuracy ranging from 83% to 100% [33].

A substantial number of tools created by other researchers have been examined, and it is evident that most of these models employ only a single form of input. It is noteworthy that there are a few instances in which two models with such divergent objectives are integrated. Furthermore, the scope of these studies is limited to two possible outcomes: a healthy state and the respective type of carcinoma.

The objective of this study is to develop a practical, low-cost, AI-based decision-support tool that integrates clinical test data (blood/stool) and, if needed, colonoscopy images to help reduce misdiagnosis and improve early detection of colorectal cancer for clinicians. The tool is composed of two components—one for blood and fecal matter analysis and the other for colonoscopy analysis. The initial model is developed using an ANN and employs a rudimentary architecture with “feed-forward” capabilities. Consequently, the obtained model can be utilized on machines with reduced computational power. The second model is created with the aid of CNN. The model under consideration features a simple structure with a limited number of convolutional and pooling layers. The utilization of these two distinct network types is predicated on the unique characteristics inherent to each. ANN is particularly well-suited for nonlinear models and can accommodate a substantial number of input variables. CNN employs image inputs and has the capacity to discern similarities among those images. In the subsequent sections of this paper, a comparative analysis and discussion of both models is presented. Furthermore, a GUI is developed that utilizes the models generated by the networks. The objective of the GUI is to enhance the accessibility of the prototype, thereby facilitating its utilization by individuals with varying degrees of familiarity, or even complete inexperience, with the MATLAB 2020a and 2022a programming language.

2. Materials and Methods

The most effective approach for achieving the objective of this study is to devise a model for each type of input. The initial instrument employed is an ANN model, whereas for the images, a CNN model is utilized. Subsequently, the models are invoked through the utilization of a user interface.

The training is facilitated by the utilization of two computers. The initial computer is equipped with a 9th generation Intel i7 processor, 32 GB of RAM, and an NVIDIA 2070 GPU (8 GB, 1770 MHz). In contrast, the secondary computer is powered by a 4th generation Intel i7 processor, 8 GB of RAM, and an NVIDIA 860 GPU. The initial computer possesses superior computational capabilities, though it may not be as readily available as the second computer. Both computers are utilized for the training of the networks and the measurement of the duration of the training process. Despite the elevated training times observed in low-end computers, the acquisition of high-performance models remains a possibility. The testing and training of these models can be accomplished without the necessity of access to expensive or specialized infrastructure.

MATLAB 2020a and 2022a is the software used for training the models and creating the GUI due to it having all the necessary tools.

2.1. The Datasets

The first step in training any type of network is to search and declare the dataset in the chosen software. For CRC, usually three main diagnosis methods are used: blood and stool testing, colonoscopy imaging, and histopathology imaging. In the current paper, colonoscopy and the data from blood and stool testing are considered.

Firstly, the numerical dataset has 36 input parameters, out of which 34 are from stool and blood sampling. The other two input parameters are age and medical history. A similar dataset from study [34] is used. The initial dataset for this study contains data from 300 patients. Out of this dataset, 100 patients are reserved for further testing, and the rest are augmented to 600 with the random feature perturbation method. The augmented dataset has been checked for clustering, data redundancy, and class imbalance, ensuring an appropriate number of patients with higher or lower cancer probability scores. In

Table 1. Class distribution table.

Dataset	Total Patients	Healthy Patients	CRC Confirmed Patients
Initial	300	170	130
Augmented	700 (600 + 100)	376 (330 + 46)	324 (270 + 54)

a class distribution table is presented.

The data are distributed into 66% for training and validation and 33% for testing. The 200 data used for training is augmented to 600, which is three times the data used for training. This value is chosen to avoid overfitting synthetic patterns and to maintain biological plausibility. The ADASYN and SMOTE methods are not considered due to having a balanced dataset. Before creating and training the neural network, these parameters need to be mathematically interpreted; thus,

Table 2. Input variables.

Variable	Min. Possible Value	Max. Possible Value	Min. Healthy Value	Max. Healthy Value	Importance Level
Age	0	120	Nan	Nan	Nan
Medical History	0	1	0	0	Nan
Albumin	0.5	10	3.5	5.2	2
Direct Bilirubin	0	30	0	0.52	3
Total Bilirubin	0	30	0.1	1.2	3
Creatinine	0	10	0.67	1.17	3
Alkaline Phosphatize	50	1000	98	279	2
Gamma GT	0	1000	11	50	2
Glycemia	20	500	70	115	3
GOT	0	500	0	37	2
GPT	0	500	0	40	2
Kalium	0	10	3.5	5.4	3
Total Protein	0	50	6.2	8	2
Natrium	0	500	130	145	3
Time Quick	0	50	13	17	3
IP	0	500	70	130	3
INR	0	50	0.84	1.1	3
Urea	0	500	18	48	3
Iron	0	500	65	175	1
Leucocytes	0	500	4	10	3
Basophiles	0	500	0	1	3
Neutrophiles nr.	0	10	2	7.5	3
Neutrophiles percentage	0	100	30	75	3
Eosinophiles	0	100	1	4	3
Lymphocytes	0	100	20	40	3
Monocytes	0	100	2	10	3
Hemoglobin	0	100	14	18	1
MCV	0	500	80	96	1
Erythrocytes	0	500	4.5	6.3	2
MCH	0	50	26	34	2
MCHC	0	100	31	37	2
RDW (rdw-cv)	0	100	8.5	11.5	3
RDW (rdw-sd)	0	50	35	36	3
Hematocrits	0	100	40	54	1
Thrombocytes	0	1000	140	440	1
Associated Pathology	0	10	1	10	3

is created. Table 2 helps visualize the analyzed parameters better and contains the minimum and maximum possible values as well as the healthy lower and upper values for each of the parameters. In addition, a grade of importance is explained below.

For most of the parameters, there is a minimum and maximum possible and healthy value. The two exceptions are age and medical history. For age, an argument could be made for minimum and maximum values, since there are average values for the youngest and oldest patients with CRC diagnoses. However, these values are highly variable based on many factors, and a non-applicable (noted as Nan in the table) value is considered more suitable for this research. The same goes for medical history since it can only be assigned the value of 1 or 0.

For each of the input variables, the importance level is granted depending on how much more importance is held by the variable. Moreover, there is a degree of severity for each variable, which is graded from 0 to 3 and is as follows:

• Means that the parameter is the healthy interval.
• Means that the parameter is slightly outside the interval 1.3 if above and 0.7 if below.
• Means that the parameter is outside the interval by 1.3–1.6 if above or 0.7–0.4 if below.
• Means the parameter is outside the interval by a larger than 1.6 if above or smaller than 0.4 if below.

After the matrix is built, an output is computed using the following rule. Depending on the matrix and how relevant that parameter is. To determine this, an experienced gastroenterologist gave his insights, and with the help of his expertise, the values were chosen. There are three stages of importance for these variables. There are 5 inputs for the first one, 9 for the second, and 19 for the third. The output percentage is computed by considering the value from the matrix and in which of these 3 categories the parameter is. The intent is to create an assistive tool to replicate and perhaps scale expert judgment, rather than as a predictive model for direct clinical diagnosis; as such, the output is set to values from 0 to 90% (a small chance is considered if some type of mistake is made or there is an off chance that some other disease causes the symptoms). Depending on the age and if the patient has had a member of their family with CRC, the sum of the 34 test result variables is modified depending on the case. It is multiplied by 1.5 if there are reports of the disease in the family. Depending on the age of the patient, the value is multiplied by 0.5–1.5 from young to old. The percentage at which it is recommended to do the colonoscopy would be around 20%, just to be sure, since early detection gives the highest chance of survival, but at around 35% the doctor should perform it.

There are 3 levels of importance for each of the variables. The ones in the third category, which have less impact on the score, are not going to be analyzed in more detail. In

Table 3. The impact of the main variable.

Variable Name	Impact of the Variable
Albumin	It can be used to track the nutritional and systematic inflammatory status of cancer patients. It has been found to be an independent prognostic in colorectal cancer, lung cancer, lymphoma, and even breast cancer. Usually, symptoms associated with high levels of albumin are dehydration and diarrhea.
Alkaline Phosphatase (ALP)	It is an enzyme associated with hepatobiliary diseases. Several reports show that ALP is used as a prognostic for colon, lung, and liver cancer (or other liver-associated diseases).
Gamma GT	It is known as gamma-glutamyl transferase, and it has been shown to have an increase in cancer patients. Usually a marker for diabetes, cardiovascular, or kidney-associated diseases, but recently it has been proven to be related to different cancer types as well.
GOT	An enzyme found in the liver and heart cells is released in the blood when either of these are damaged. A high value can be associated with cancer.
GPT	An enzyme that is found in the liver and some other tissues. If it is found in the blood, it may be a sign of liver damage or cancer. People who present high values for the GPT are sent to get evaluated by gastroenterologists.
Total Protein	Parameters are usually measured to detect the amount of protein in a person’s blood. These tests can be used to diagnose several conditions like cancer and kidney or liver diseases.
Iron	Excess iron, which cannot be assimilated by metabolism, can be associated with cancer. The most relevant one is leukemia, but in the case of colorectal cancer, there is a deficiency of iron and anemia.
Hemoglobin	The level of hemoglobin in a cancer patient is usually found in a declining state; it can start even 4 years before the diagnosis of the disease. A low hemoglobin level is associated with anemia.
MCV	It is used to measure the average red blood cell in a person. It can be used to detect B12 deficiency, but if the value is high, it can be associated with cancer.
Erythrocytes	Red blood cells in low numbers can be associated with cancer.
MCH	Mean corpuscular hemoglobin is the average amount of hemoglobin found in the blood cells.
MCHC	Mean corpuscular hemoglobin concentration is the average amount of hemoglobin found in a group of red blood cells.
Hematocrits	It is the volume percentage of red blood cells in blood.
Thrombocytes	Also known as platelets, they are the cells in the blood that form the clots. An elevated platelet count can be linked to the development of tumors.

the impact of the main input variables is presented.

Next, the colonoscopy imaging diagnosis method is considered. For this model, the Kvasir dataset [35] is considered. It has several folders with different images that could be used for the detection of colorectal cancer. The colonoscopy images from it are used for the training of CNN. There are 2 cases that are analyzed. In the first case, 500 healthy images and 500 images that were previously diagnosed as positive for CRC are used. For the second case, 500 images of ulcers are added.

Figure 1a presents a colonoscopy image of a colon positively diagnosed with CRC. In the middle of the image, the polyp can be seen as a tumor. Figure 1b is what a healthy colon looks like.

Figure 2 shows gray images created for faster training and higher resolutions.

As previously stated, the public Kvasir dataset is employed, comprising a total of 1000 images of colonoscopies. The images are divided into two categories: one for healthy images and one for images that contain polyps (patients affected by CRC). In experimental settings, a third label is incorporated for images depicting ulcerated colons. Given that dimensions constitute a significant challenge, the images undergo resizing to ensure manageability. Initially, the images are resized to 128 × 156, and subsequently to a range of larger resolutions, with the objective of preserving a substantial amount of information. Given that the issue pertains to the computational capacity, the images undergo a transformation from RGB to gray format. Consequently, the input matrix undergoes a reduction from a three-dimensional array to a one-dimensional array (the RGB format utilizes three matrices as inputs, whereas the grayscale format employs a single matrix). The benefit of this approach is that it significantly reduces the number of computations. However, it should be noted that this reduction in computations also results in a concomitant reduction in the amount of information. Even if the network is adequately trained, there remains the possibility of a model exhibiting suboptimal performance.

Images of varying resolutions are generated, commencing with 128 × 156 and culminating at 720 × 720. The duration of training is directly proportional to the dimensions of the images and the quantity of matrices employed during the training process. The utilization of larger images is imperative to ensure the preservation of essential information. The training process may require a greater investment of time or may not result in successful completion due to limitations in memory capacity. However, this process is essential for the development of a more effective model, although its impact may not be immediately apparent in performance outcomes. It is imperative to establish a specific size for the images, given that the initial layer of the network is designated as an input layer. This necessitates the provision of the number of rows and columns as parameters. The quantity of matrices is indicated by the number of gray matrices or RGB matrices. An error is indicated for 720 × 720 due to the size being too large.

Subsequently, the images are rotated at various angles, including 90, 180, and 270 degrees, with the objective of having different perspectives of the images, being harder to rotate automatically in the GUI. Given that a subset of these images is utilized for the purpose of validation, and that the performance is evaluated based on the validation images, it is imperative to ensure that enough of these images are available. However, it should be noted that 100 healthy images and 100 polyp images are stored separately for the purpose of testing after the training procedure is completed. The remaining 800 images from the two labels are then subjected to a rotation at the previously specified angles. This incorporation results in an augmentation of the image collection, bringing the total to 3200. The primary challenge associated with this approach pertains to the extended duration required for network training. The images in question are selected at random from the image pool. The same procedure is applied to the case with three possible outputs. In this particular instance, the initial dataset consists of 1500 instances, with 500 instances assigned to each label. In accordance with the protocol, 100 images are retrieved from each label for the testing phase. The remaining images undergo a rotational process, akin to the previous case, ensuring a balanced distribution across all labels. The augmented dataset comprises 4800 images in total, with 1600 images assigned to each label. . In

Table 4. Class distribution table used for training.

Case	Dataset	Total Images	Healthy	Polyp	Ulcer
3 possible labels	Initial	1200	400	400	400
	Augmented	4800	1600	1600	1600
2 possible labels	Initial	800	400	400	0
	Augmented	3200	1600	1600	0

the class distribution is presented.

2.2. ANN-Based Tool

ANN is a powerful tool when talking about AI-based modeling. The inspiration for these devices stems from the functionality of the human brain. The axons can be conceptualized as the nodes, and the dendrites and terminals can be regarded as the weights and biases, respectively.

The neural network employs the previously mentioned map values that have been provided as an input. The matrix is finalized with the incorporation of age and family history values. During the training process of the network, certain parameters are subject to modification. These include the number of neurons and the number of hidden layers. The Levenberg–Marquardt (L-M) training function is the most commonly employed, though the Quasi-Newton and Scaled Conjugate Gradient methods have been utilized in select experiments. It has been demonstrated that L-M is superior to the other options [7]. The feedforward architecture is utilized for the structures. The activation function is implemented using the sigmoid function for the hidden layers, while the default option is employed for the output layer, namely the linear function. The stopping values are set to 10,000 epochs, the gradient is set to 10⁻³⁵, 30 validation checks are performed, and the maximum MU is set to 10³⁵. These values were modified to increase the probability of obtaining a more suitable model.

As illustrated in Figure 3, the percentage values that were calculated using the previously outlined formula are displayed. These values will be the output targets when training the neural network. The median value of 35% is traced to facilitate the visualization of patients who fall below or above this threshold value. The maximum value is also considered, and it is only reached by the last patient. This is used to scale the values.

In Figure 4, the network architecture is presented. There are 2 hidden layers with 35 neurons on each. The activation function is the sigmoid function as presented before, and the feedforward structure is used.

As illustrated in Figure 5, the network’s training procedure is outlined as follows. The stopping parameters are presented and traced. As illustrated in the image, the maximum MU parameter was the factor that led to the cessation of the training process. Typically, the training procedure requires between 10 and 30 s, as the network is relatively uncomplicated.

The performance of the models is compared with the help of the MSE. This index quantifies the extent to which the predicted values deviate from the experimental data.

The ANN-based model is employed to predict the likelihood of CRC presence in patients, yet it lacks the capacity to discern the location of polyps or ascertain the presence of other diseases with comparable biological characteristics. To enhance the precision of the diagnostic process, a second model has been developed that utilizes CNN.

2.3. CNN-Based Tool

The training of the CNN utilized the following architecture. Initially, three convolutional layers and three pooling layers are employed, with the incorporation of normalization and ReLU (rectified linear unit) layers. The first convolutional layer is characterized by a 3-by-3 filter size and eight filters, with the padding option set to ‘same.’ The second convolutional layer maintains the same filter size, but the number of filters is augmented to 16. The final convolutional layer contains 16 filters, each set to the same dimensions of 3 by 3. Maximum pooling is employed in all pooling layers, with dimensions set at 2 by 2.

After this segment, a fully connected layer is implemented, accompanied by the integration of classification and softmax layers. The configuration of these layers is determined by the number of outputs utilized, with 2 or 3, respectively. The dimensions of the images are required as inputs for the input layer. The number of labels (outputs) is required for the fully connected layer.

• ReLU layer: The role of the ReLU layers is to linearize the computation increase, which without these layers would grow exponentially.
• Normalization layer: The normalization layer is used to normalize the output of the previous layers; it also allows the network to learn more independently from layer to layer.
• Convolutional layer: The convolutional layer is the main building block of any CNN, and it is usually a set of filters that are determined while the network is trained. Due to the convolutional layers creating feature maps, with each convolutional layer the dimensions of the maps increase.
• Pooling layer: To reduce the dimensions of the feature maps, pooling layers are used.
• A fully connected layer is a feed-forward neural network that has as its input the flattened output of the final convolutional and pooling layer.
• Classification layer: usually used to compute the cross-entropy loss for classification tasks.
• Softmax layer: used to predict a multinomial probability distribution for multiclass classification problems where two or more class labels exist.

After the architecture is set, some parameters need to be given, such as the initial learning rate, set at 0.01, and the number of epochs, set at 100. The data are then shuffled around after every epoch, mainly to improve accuracy. Then, the validation data are given; for every type of NN, there is a requirement of having training data and validation data. Lastly, validation is performed every 30 iterations.

3. Results

After obtaining the models for each of the two networks, the one with the best performance is chosen from each type of network and implemented in the GUI.

3.1. The ANN Model Results

After the previously mentioned training parameters are set, the following results have been obtained.

As demonstrated in

Table 5. Results for the ANN-based models.

Training Function	Hidden Layers	Number of Neurons	Performance (Mean Squared Error)	Training Time
Levenberg–Marquardt	2	35	7.38	9 s
Levenberg–Marquardt	7	35	19.09	3 min 46 s
Levenberg–Marquardt	7	20	13.3	26 s
Levenberg–Marquardt	2	20	11.64	2 s
Quasi-Newton	2	20	15.76	3 s
Quasi-Newton	7	20	17.95	7 s
Quasi-Newton	7	35	14.4	15 s
Quasi-Newton	2	35	19.78	7 s
Scaled Conjugate Gradient	2	35	13.07	1 s
Scaled Conjugate Gradient	7	35	17.22	2 s
Scaled Conjugate Gradient	7	20	15.90	2 s
Scaled Conjugate Gradient	2	20	13.96	1 s

, the findings of training the network with varying training parameters are presented. The Levenberg–Marquardt algorithm demonstrated the optimal performance outcomes. The Quasi-Newton method demonstrated commendable performance and expeditious processing times in numerous instances. However, under certain circumstances, the MATLAB environment exhibited instability, necessitating a restart of the process. Regarding the issue of speed, the training algorithm that has been demonstrated to be the most effective in this case is the Scaled Conjugate Gradient algorithm. Even in scenarios involving a greater number of neurons and hidden layers, there was no observed change in the training time.

As depicted in Figure 6, the performance of the model obtained with the L-M training function with two hidden layers and 35 neurons on each layer is demonstrated. The model’s intricacy, as evidenced by its higher complexity compared to the other results, is noteworthy. However, it is notable that the model exhibits an optimal performance with an MSE of 7.38, which is a testament to its efficacy. Several studies have obtained smaller MSE values; however, it should be noted that the amount of data in these cases is comparatively limited. This performance type is characterized by its additive nature, which signifies that the value increases in proportion to the quantity of data utilized. Given the total of 600 values utilized and the scale at which they are measured, the performance obtained is satisfactory. In the training of these networks, a 5-fold cross-validation procedure is employed for each instance. Furthermore, performance metrics are documented for the 100 patients who have been reserved exclusively for the model that exhibited the optimal performance. The model’s accuracy has been determined to be 95%. The model’s specificity is 95.65%, and its sensitivity is 94.44%. The confusion matrix for the obtained model is illustrated in Figure 7.

3.2. The CNN Model Results

The architecture of the CNN has already been discussed, and the only changes occurring throughout the experiments are affecting the input resolution of the images. Firstly, the resolution of the images has increased from 128-by-156 to 640-by-640. These values must be specified when declaring the training parameters. Secondly, a third value for the resolution of the images is employed to ascertain whether the images are grayscale or RGB. The training parameters have been established. After the procedure, the following window will appear.

Figure 8 presents information about the trained CNN, focusing on accuracy and training time. The cycle length is user-defined (20 epochs with 23 iterations per epoch). Performance metrics include accuracy, measured as the percentage of successful predictions. For example, with 100 images and 97% accuracy, the model correctly predicts the outcome for 97 images.

Next, the results of CNN-based models are presented.

Table 6. Training results for the CNN-based models with two outputs.

Case	Resolution	Performance	Training Time (min:s)
Kvasir gray	128 × 156	91%	0:35
Kvasir gray rotated	128 × 156	92.5%	2:01
Kvasir rgb	128 × 156	93%	3:07
Kvasir gray rotated	256 × 256	96.7%	5:52
Kvasir rgb rotated	256 × 256	97.1%	9:16
Kvasir rgb rotated	512 × 512	96.9%	46:42
Kvasir rgb rotated	640 × 640	94.5%	120:29

shows the performance of models with two outputs. The cases analyzed include gray and RGB images at various resolutions. Performance, measured as the percentage of images correctly labeled, is the more important characteristic, while training time is less important. Training time increases with higher image resolution and with RGB images instead of gray. This is due to increased computations. In RGB images, the number of matrices defining the image increases from 1 to 3. The best performance, 97.1%, is obtained at a resolution of 256 by 256.

As illustrated in Figure 9, the progression of accuracy is demonstrated. In accordance with the principles that govern any training modality, the initial parameters are selected at random, and subsequent refinements to the performance can be made. In the initial 30 epochs, the error value for the modified validation images exhibited greater variability. However, in subsequent epochs, the error remained relatively constant and did not undergo significant fluctuations. In the context of experimentation, the number of epochs remains constant. However, it is possible to attain equivalent performance levels while utilizing a reduced number of epochs.

For a better evaluation of the obtained models, a 10-fold cross-validation is used with stratified splitting to avoid class imbalance. The images are divided into 10 folds, using one for testing and the rest for training. This technique is used to have a better generalization estimate. A slight increase in performance has been observed when compared to models that have not employed the 10-fold cross-validation. All the models in this study used the same 10-fold cross-validation for the obtained CNN models.

In Figure 10, the confusion matrix for the trained CNN is presented. Healthy and polyp-labeled present 99 and 97 predicted values out of 100. The confusion matrix is created using only the 200 images reserved for testing from the initial dataset.

Since the best results have been obtained for the 256-by-256 resolution, it has been chosen for the CNN with three possible outputs.

The CNN model is trained with a new output for detecting ulcers. As performed before, 500 images are added, rotated, and resized for training. This improves patient diagnosis. The training process is illustrated in Figure 11.

The validation accuracy for the second CNN is 83%, with a training time of 4 min for 700 iterations. The validation percentage has worsened since the value dropped from 97.2% to 83%, but this is expected when introducing a new label to the existing ones.

As evidenced by the data presented in Figure 12, the accuracy of the trained CNN is 82%, requiring a mere three minutes of training. The two new models have been thoroughly trained in the optimal performance-yielding parameters and have been utilized with the GUI.

As illustrated in Figure 13, the confusion matrix of the 256 × 256 CNN is displayed. A comparison is made between the predicted outputs and the actual outputs. The healthy label has been demonstrated to be predicted 97 times out of a possible 100 values. This results in an accuracy rate of 97% for the label. For the polyp label, this value is reduced to 80 out of 100, indicating a decrease in accuracy to 80%. Finally, the ulcer label indicates a value of 87 from a possible range of 100, which corresponds to an accuracy of 88%. The confusion matrix is created using the reserved 300 images from the initial dataset.

CNN model performance is comparable to literature results. The model with two labels had an accuracy of 97.2%, but higher values require an accuracy or complexity increase. The architecture is kept simple, and the results are good. The model with three labels has an accuracy of 88%, but it is still a good model. The ability to differentiate between CRC and ulcers is useful and is not commonly analyzed.

A separate study was conducted to ascertain the efficacy of transfer learning when employing the “squeezenet” pre-trained architecture from the deep learning toolbox. The 256-by-256 resolution was selected for its superior performance in this study. Following the establishment of the input and output layers and the training of the network, the subsequent confusion matrix was obtained for the 2-output model.

As illustrated in Figure 14, the accuracy of the pretrained network is 99.5%. The data utilized in this study are consistent with that employed in the preceding cases, with 3200 instances allocated for training and 200 instances designated for testing.

A similar procedure is executed for the 3-output model as well, yielding results that are comparable to those obtained from the aforementioned procedure. Figure 15 provides a visual representation of the outcomes observed. The confusion matrix for this model demonstrates an average value of 93%, with 84% accuracy for a single case. In this particular instance, the training images are the 4800 images that were utilized in the preceding cases. Concomitantly, the 300 utilized in the preceding cases are employed in this study as well.

The pretrained models employing the “squeezenet” architecture demonstrate superior performance metrics in comparison to the previously trained networks examined in this study. This enhancement in performance is accompanied by a more intricate architecture.

3.3. The GUI

The GUI is composed of two panels—one for the model created with ANN and one for the model created with CNN. The ANN panel comprises two buttons: one for uploading data and one for computing probability. Upon engaging the “load data” function, a novel browser window will initiate, within which the user may navigate to the file containing the patient’s data. Following the selection of the desired file, the data are loaded into the interface in the form of a table, thereby rendering each value visible. Subsequently, the “compute probability” button can be engaged, which should result in the appearance of the value. In the absence of loaded data, an error message is displayed, prompting the user to upload data. As illustrated in Figure 16, the ANN panel is presented after a successful execution.

The subsequent segment of the program is the presentation of the CNN panel. As with the ANN panel, the device is equipped with two buttons—one for uploading the image and one for image analysis. Upon pressing the “load image” button, a window will appear, and the user will be prompted to locate the desired image. Following the loading of the image, the subsequent button must be engaged to ascertain whether the image contains polyps or not, or if the colonoscopy image manifests as ulcers. In the event that the buttons are engaged in an order that is not correct, a message is displayed to inform the user that an image must be uploaded. As illustrated in Figure 17, the execution of the GUI was successful.

The tool works by allowing users to choose what needs to be analyzed. Blood and fecal tests can be run on the ANN panel, while colonoscopy images can be loaded into the CNN panel. The two panels can be used on the same patient for a more precise diagnosis or on different patients. The only necessity is to upload the correct type of data onto the panels.

This GUI unites the two networks to provide a more thorough diagnosis, analyzing blood test results and colonoscopy imaging. These are the two main ways doctors make decisions. Literature usually only discusses one of these diagnoses, and models are obtained for either image inputs or test results. Having two sources for diagnosis can be beneficial in situations where the results are not straightforward. For example, the ANN tool could give a CRC probability of around 50%, but when the images are loaded, no polyps are detected. In that situation, there might be other medical conditions causing the unhealthy values of the biological parameters.

A benchmark was conducted to assess the memory utilization of the GUI. The results indicated that a maximum of 700 megabytes of memory were consumed on both workstations. This is attributed to the fact that the GUI does not involve the generation of intricate plots, the process of rendering, or the loading of voluminous datasets. The benchmark encompasses the initiation of processes, the upload of data and images, and the execution of ANNs and CNNs. At the higher end of the spectrum, each action was executed in seconds. The highest execution time was recorded at the start-up stage, with an average duration of five seconds, while the entire benchmark test took approximately one minute. The preliminary examination of the lower-end workstation revealed that the startup process takes an average of 15 s. The subsequent benchmark assessment indicated that the process takes approximately one and a half minutes to complete.

The installation of the tool on a computer could be executed without the requirement of a high computational power. The gastroenterologist is able to enter the patient’s test results into the application. The subsequent course of action can be determined by referring to the result. In the event that the CRC probability is sufficiently elevated, the physician will proceed with the colonoscopy. Following the procedure, the physician is able to upload the images captured during the colonoscopy into the designated software tool. Subsequently, by pressing a designated button, the physician can proceed with providing a diagnosis to the patient.

4. Discussion

Firstly, a discussion of the ANN model results is warranted. A thorough examination of the available literature reveals that there are three training algorithms that have been analyzed. The L-M algorithm has been shown to yield optimal results, demonstrating both superior performance and expedited training times. It is regarded as the superior and most versatile training algorithm, having been employed in numerous models from diverse fields. In determining the most suitable architectural design, two fundamental factors must be taken into consideration—simplicity and performance. In the domain of mathematical modeling, the quality of a model is contingent upon the number of hidden layers and neurons employed. It is imperative to note that the efficacy of the model is subject to limitations, which are determined by these parameters. Following a series of empirical trials, it was determined that the optimal model exhibited an MSE of 7.38 for two hidden layers comprising 35 neurons. The system’s simplicity is attributable to its feedforward structure, which is widely regarded as one of the most elementary structures. A comparison of the results with those obtained from models derived from literary sources reveals no statistically significant differences. In the study [7], many CRC detection models are discussed. The best ANN-based model has 93% accuracy and uses SELDI for proteomic analysis, but it is different from the data in this paper. The most similar approach had 74% accuracy and used the TNM staging system. However, the model has its limitations. Due to using expert-derived scores, the obtained model may have limited generalizability and objectivity.

The CNN model is chosen based on multiple factors. It could be the model with the smallest resolution, given its small training time and low computation requirements. The accuracy is in the 90% range, but using a small resolution can lead to information loss. The GUI rescales the input image to a similar resolution as the training image, reducing the loss of information. A medium resolution is selected to minimize information loss and computational demands. The accuracy is comparable to results from other studies. In the study [7], several CNN-based models are presented with a wide range of values for accuracy. Simple structures have values from 75% to 90%. Complex structures reach 99%. This study balances complexity and performance and explores additional diagnoses with the output of ulcers. Adding labels and images reduces accuracy, showing that many labels for diseases could harm a CNN model.

The GUI uses the models to make the tool user-friendly for medical use. Many studies compare neural network-based models, but few use multiple models for diagnosis. This study’s contribution is a more complex tool for thorough diagnosis using both types of input variables. Nevertheless, the data originate from two distinct sources that are not interconnected: the Kvasir dataset, utilized for the CNN model, and an augmented novel dataset, employed for the ANN model. It has been demonstrated that the two datasets can be utilized to train adequate models with their respective datasets. However, it cannot be concluded that better performance could be obtained by using a combination of both models for diagnosis.

In the context of future developments, the potential exists for the creation of a CNN-based tool that would have the capability of coloring affected zones of CRC or of establishing a bounding box around these zones with the objective of highlighting them. Firstly, the utilization of this technology has the potential to reduce the time required for medical personnel to complete a diagnosis. Furthermore, it may serve to illuminate areas that might not otherwise be given full consideration. Secondly, the primary objective of these instruments is to mitigate the likelihood of misdiagnosis. Finally, the visual nature of the image is more appealing than that of a text message.

The GUI could also be altered to incorporate additional utilities. For instance, buttons that facilitate the resetting of the GUI to remove superfluous data from the interface. In certain instances, the information may not be necessary, which can result in challenges or issues for the user. Another potential enhancement would be the incorporation of a “help” button that would provide a detailed explanation of the operational mechanisms of the interface. Typically, a window providing a description of the buttons’ functions or the steps required to utilize an interface is employed. The other possibility is to search the documentation for the necessary information. The utilization of statistical tools can also be advantageous in this context. The ability to ascertain the percentage of patients who have been diagnosed at the end of a day or shift provides valuable insights for research and statistical purposes, benefiting the staff.

An alternative approach for enhancing this instrument entails the establishment of libraries that employ a variety of models and possess the capacity to select between different organs or diseases. The viability of such a project is contingent upon the amount of time allocated to its implementation.

A subsequent study could be conducted within a clinical setting to evaluate the performance of the application in a real-world context. The results could potentially reveal methods to enhance the prototype or substantiate the assertions through statistical data. Such statements are indicative of a reduction in misdiagnosis or an improvement in early detection rates.

5. Conclusions

In summary, the goals of obtaining high-performance models with simple network architectures have been achieved. For the ANN-based model, a 7.38 value for MSE, considering the high amount of data for which it is computed, is a high enough accuracy to be considered a good model. However, the simpler structure could be advantageous when using the application on a low-performance computer. The same could be said about the CNN model. When employing a simple architecture, the accuracy at training is 97% for the two-output models and 83% for the three-output model. Despite the presentation of lower accuracy, the model unveils a distinct diagnosis that is seldom analyzed in the majority of studies. Furthermore, as the architecture becomes more intricate, the precision of the two-output model elevates to 99.5%, while the three-output model attains 93%. In addition, GUI has a very important role in uniting the two models and delivering a better diagnosis. The application being user-friendly makes it easier for doctors or medical teams to use it.

The application can be further developed into a library with several types of organs from which the doctors could choose to aid them in diagnosing patients. Another option could be to add more types of diseases, like irritable bowel syndrome, and retrain the networks with more data.