Efficient-gastro: optimized EfficientNet model for the detection of gastrointestinal disorders using transfer learning and wireless capsule endoscopy images

Limitations and research gap

Advancements are being made in the classification of gastrointestinal disorders, despite the continued limitations associated with early-stage diagnosis, the scarcity of training datasets, and the need for high-quality imaging. Certain elements that contribute to such concerns are the complexity of the model design and the substantial processing demands. However, improvements to datasets, streamlining of model construction, and the development of methods accommodating varying degrees of image quality remain potential. Many studies combined different deep models that increased the computational time and were less accurate. The use of raw images for feature extraction has been observed to reduce classification accuracy. There is no proper augmentation techniques used to enhance the dataset and reduce the overfitting problems in the literature. The objective is to construct optimized deep learning models capable of early disease detection, potentially leading to improved patient outcomes and increased survival rates.

Research questions

This work aims to address three main research questions, and the answers are provided in the results and discussion section.

RQ1: What is the impact of choosing and integrating data augmentation approaches on the performance of the optimized EfficientNet model for gastrointestinal disease detection?

RQ2: How effective is the model in terms of complexity and processing speed?

RQ3:: Is the model robust to different datasets.

Research motivation

The gastrointestinal system is essential for the identification and treatment of a number of diseases, such as stomach and esophageal malignancies. The gastrointestinal images in the Kvasir collection represent an important area of study that has the potential to greatly enhance medical practices and healthcare facilities worldwide. For effective treatment and ongoing surveillance, the diagnosis of gastrointestinal cancers—in particular, colorectal cancer—is essential. However, medical professionals’ proficiency in detecting a condition may have an impact on the precision of the diagnosis. Enhancing the evaluation and detection of gastrointestinal disorders using automated diagnostics might result in higher productivity and better use of available resources. The availability of extensive medical databases and technological advancements in healthcare procedures has greatly improved the diagnosis and treatment of a wide range of disorders. The proficiency of endoscopists and factors related to the gastrointestinal tract compromise detection accuracy. It is thought that this investigation’s findings will be very important in lowering the number of gastrointestinal cancer cases and fatalities.

Research contributions

The research contributed the following points:

• We proposed an EfficientNet method with fine-tuned Conv2D layers and Dropout layers to accurately and efficiently detect and classify gastrointestinal disorders, utilizing two datasets that are available to the public. One dataset has eight diseases, and the second dataset has four gastrointestinal-related diseases. The main aim of this method is to build a lightweight, reliable, and less parameter-based approach that makes detection easier.

• The collected dataset is then augmented to enhance the number of training images for deep learning models and preprocessed to enhance its quality.

• The pretrained deep learning models are used for the Detection of gastrointestinal tract disorders, and previous handcrafted methods are compared with the proposed approach. Accuracy, precision, recall, F1 measure score, AUC, loss, and true and false predictions are all used to assess the significance. Also, a detailed analysis is performed.

• To address the data overfitting problem and demonstrate the model’s effectiveness, this study utilized cross-validation experiments. The accuracy of the proposed method outperforms that of other state-of-the-art methods.

The subsequent sections of the article are structured as follows. ‘Related work’ offers a summary of the research that has been conducted on the detection of gastrointestinal disorders through the use of wireless endoscopic data sets. In ‘Proposed methodology’, the methodology that is offered is described in detail. This technique involves the preprocessing of the dataset, dataset augmentation, the implementation of the proposed method, and the evaluation of the effectiveness of the proposed method using performance metrics. Results and discussion are presented in ‘Results and discussion’. The conclusion and recommendations for further research are presented in ‘Conclusion and future work’.

Preprocessing and data augmentation

The main theme of pre-processing is to enhance the quality of an image in order to facilitate more effective evaluation. Preprocessing techniques are employed to minimize undesirable distortions and enhance particular features that are crucial for the desired result. Preprocessing is essential to resize the original images in the dataset into similar or fixed sizes for the convenience of deep models because the original images in the dataset are in various shapes and dimensions (Mujahid et al., 2022). We resized the images into equal sizes using the Python module of computer vision. Additionally, image preprocessing can reduce model training and inference times. when the images being used are especially large, reducing their size will significantly decrease model time for training without substantially affecting the model’s performance.

Figure 2 shows the samples of gastrointestinal tacts. Augmentation is the process of creating new samples from existing samples in a dataset to enhance the dataset, and preprocessing is the process of adjusting existing images to satisfy certain requirements. The Image Augmentor Python package is deployed to enhance the data class images during image augmentation process. Algorithm 1 details the process of data enhancement. The rotation is the first stage in a series of phases that make up the data augmentation process. The zooming, flip_left_right, and flip_top_bottom operations come next. Last but not least, we employ a sample rate and the random_distortion function. In order to prevent model overfitting for data from the majority class, new images are produced to balance the dataset. Splitting the augmented dataset into 90:10 ratio is illustrated in Table 2.

 
_______________________ 
Algorithm 1 Augmentation._________________________________________________________________ 
Input:_ Training Images 
Output:_ Augmented Dataset 
import Augmentor 
Start: 
  1:  Aug_pipeline ← Augmentor.Pipeline(”images_path”) 
  2:  Aug_pipeline ← (Zoom_probability = 30%,percentage_area = 40%) 
  3:  Aug_pipeline   ←  (Rotate_robability   =    30%,Max_left_rotation   = 
     5,Max_right_rotation = 5) 
  4:  Aug_pipeline ← Flip_Left_Right = 30%) 
  5:  Aug_pipeline ← Flip_Top_Bottom = 30%) 
  6:  Aug_pipeline  ← Random_distortionprobability  =   40%,Gridwidth  = 
     2,Gridheight = 2,magnitude = 7) 
End__________________________________________________________________________________________________    

Proposed method

EfficientNet-B1 is a part of the family of EfficientNet and was created by the researchers at Google in 2019. The lowest computational time, efficiency, model size, and high perfromance and effective architecture are the power of this EfficientNet family. The proposed methodology incorporates the utilization of a CNN architecture and a scaling method that employs a compound coefficient approach. This approach ensures the consistent scaling of all parameters related to depth, width, and resolution. The EfficientNet scaling strategy involves a systematic increase in the width, depth, and resolution of the network using a predefined set of scaling coefficients. This differs from the conventional approach, which often scales these elements independently, potentially resulting in a less robust model. The compound scaling strategy is justified by the requirement to accommodate larger input images by incorporating extra layers. This allows for an improved receptive area within the network, enabling the identification of more significant elements over the entire image. Additionally, the inclusion of more channels further enhances the network’s ability to detect key features. EfficientNet provides high accuracy and is specially adopted in such a case where there’s limited memory and resources. It works efficiently and achieves outclass results in computer vision and classification-related tasks. Figure 3 displays the proposed method’s architecture.

The authors used the EfficientNet architecture with the transfer learning process because building an architecture from the initial stages is more challenging and time-consuming. Transfer learning is a technique in which learned knowledge is again used for another task by fine-tuning the parameters and modifying the additional layers. To minimize the total parameters and overfitting, the authors implemented a global pooling average layer to the EfficientNet model. A number of dense layers with dropout layers and relu activation functions was also used. To avoid overfitting, a dropout rate of 35% was chosen at random. A softmax activation function was used to one output dense layer with four output units for multiclass classification in dataset 1 and eight output units in dataset 2 build the proposed automated model. Also, two Gaussian noise layers with a 25% rate each were added.

Throughout the experiments, open-source software and libraries are utilized. To speed up the training, the authors used Google Colab Notebook with GPU run time. In simple terms, the software utilized has no license constraints since it is both open source and free to use.

 
_______________________________________________________________________________________________________ 
Algorithm 2 Algorithm for the Proposed method.____________________________________ 
base_model = EfficientNet 
weights = ’Imagenet’ 
input_shape = (224 * 224 *3) 
Mo = model 
Start: 
  1:  Mo = Sequential() 
  2:  Mo ← base_model 
 3:  Mo ← Conv2D(128,kernelsize = (3,3),activation =′ relu′) 
  4:  Mo ← Conv2D(512,kernelsize = (3,3),activation =′ relu′) 
  5:  Mo ← GaussianNoise() 
  6:  Mo ← GlobalAveragePooling2D() 
  7:  Mo ← Dense(512,activation =′ relu′) 
  8:  Mo ← BatchNormalization() 
  9:  Mo ← Dropout() 
10:  Mo ← Dense(512,activation =′ relu′) 
11:  Mo ← BatchNormalization() 
12:  Mo ← GaussianNoise() 
13:  Mo ← Dense(8,activation =′ softmax′) 
End__________________________________________________________________________________________________    

Dropout regularization is a technique that is similar to training multiple neural network architectures simultaneously. A subset of layer results are arbitrarily “dropped out” or neglected during training. Any change made during training utilizes a distinct “view” of the layers in its final state. Dropout layers are very crucial to in the training process for deep learning models as they prevents from overfitting the data. The first set of samples for training has an abnormally high level of effect on the learning process and thus prevents the learning of features that appear only in later batches. Batch normalization is a method of supervised learning that normalizes the intermediate outputs of a neural network by converting them into a standard structure. This process is also known as normalizing. This “resets” the distribution order of the output from the layer previous it so that the layer above it can handle it more effectively. After removing the mean batch and scaling by the standard deviation, the output of the preceding activation layer is transformed by batch normalization. This makes a deep learning network more stable. Table 3 represent model architecture and their parameter values for gastrointestinal disease detection.

Experimental setup

The efficacy of the proposed approach was assessed by employing Google Colab and two datasets sourced from the Kaggle repository. The equipment that was utilized comprised an Intel Core i7-7500U central processing unit, 32 GB of random-access memory, 24 GB of video memory capacity, and an NVIDIA GeForce RTX 3080 graphics card. TensorFlow and PyTorch were employed to enhance computing efficiency. The selection of Python 3.10 was based on its comprehensive library and adaptable nature. The study examined augmentation and preprocessing techniques and compared the model’s accuracy, training duration, and detection speed with previous studies. The datasets were partitioned into two distinct groups for analysis: 90% for the purpose of training and 10% for the purpose of testing. The study seeks to make a substantial contribution to the field of medical image processing, specifically by focusing on the identification of gastrointestinal disorders. Deep learning technology integration has the potential to enhance diagnosis accuracy, resulting in substantial treatment advantages. The timely and accurate detection of gastrointestinal problems holds great clinical importance. Table 4 presents the training configuration of proposed Model architecture.

Performance metrics

Classification problems are among the subjects that have received the greatest research attention everywhere on the globe. Use cases may be found in almost every medical setting. Identifying different writers, voices, and the content of written words. We need a metric that compares discrete classes in some way in order to make use of the fact that the output from classification models is discrete. Classification metrics evaluate the performance of a model and provide information on how accurate the classification is; however, they accomplish this in a variety of different ways (Padilla, Netto & Da Silva, 2020).The ratio of correct predictions relative to the total number of predictions, multiplied by 100, is called classification accuracy. This may be the statistic that is easiest to understand and use. The authors employed the abbreviation “_g” in conjunction with the metrics formulas, wherein “_g” represents the gastrointestinal.

Accuracy in classification models is a metric that quantifies the proportion of accurately predicted instances. It is determined by dividing the number of correct predictions by the total number of predictions made. (1)${\displaystyle \mathrm{Accuracy}\text{\_}\mathrm{g}=\frac{\mathrm{TP}\text{\_}\mathrm{g}+\mathrm{TN}\text{\_}\mathrm{g}}{\mathrm{TP}\text{\_}\mathrm{g}+\mathrm{TN}\text{\_}\mathrm{g}+\mathrm{FP}\text{\_}\mathrm{g}+\mathrm{FN}\text{\_}\mathrm{g}}\ast \left(100\right)}$

Precision is the capacity of a model to reliably predict positive results. The ratio of accurate positive predictions to erroneous positives determines it. High accuracy signifies the effective mitigation of false positives, particularly in contexts such as medical diagnosis. (2)${\displaystyle \mathrm{Precision}\text{\_}\mathrm{g}=\frac{\mathrm{TP}\text{\_}\mathrm{g}}{\mathrm{TP}\text{\_}\mathrm{g}+\mathrm{FP}\text{\_}\mathrm{g}}\ast \left(100\right)}$

Recall is an essential classification parameter that measures a model’s capacity to accurately identify as well as incorporate true positives. Optimizing recall is essential in medical diagnostics since the failure to detect positive instances might result in significant repercussions. (3)${\displaystyle \mathrm{Recall}\text{\_}\mathrm{g}=\frac{\mathrm{TP}\text{\_}\mathrm{g}}{\mathrm{TP}\text{\_}\mathrm{g}+\mathrm{FP}\text{\_}\mathrm{g}}\ast \left(100\right)}$ An important classification metric that balances recall and accuracy to provide a fair assessment of a model’s utility is the F1-score. It strikes a compromise between memory and accuracy, particularly when obtaining real positive instances and lowering false positives. It is frequently applied to assess the overall performance of the model. (4)${\displaystyle \mathrm{F1}-\mathrm{score}\text{\_}\mathrm{g}=2\ast \frac{\mathrm{pre}\text{\_}\mathrm{g}+\mathrm{Rec}\text{\_}\mathrm{g}}{\mathrm{Pre}\text{\_}\mathrm{g}\ast \mathrm{Rec}\text{\_}\mathrm{g}}\ast \left(100\right)}$

In evaluating the efficacy of classification models, the AUC, which represents the area under the ROC curve, is a crucial metric. An AUC value exceeding 1 indicates perfect categorization, while a value below that indicates superior discrimination. It provides a comprehensive viewpoint regarding the sensitivity and specificity of a model. (5)${\displaystyle \mathrm{AUC}\text{\_}\mathrm{g}=0.5\ast \left(\mathrm{TPR}\text{\_}\mathrm{g}+\mathrm{TNR}\text{\_}\mathrm{g}\right)\ast \left(100\right).}$

Datasets

Automated identification of multiple diseases through the use of technology such as computers is a significant yet unexplored area of research. These innovations might enhance healthcare services and improve global health care networks. However, databases with medical images are limited, which makes reproducibility and method comparison nearly impossible. In this article, we provide the Kvasir dataset, which contains images of the gastrointestinal diseases. The dataset has been arranged and annotated by medical professionals (Expert Endoscopists). Norway’s Vestre Viken Health Trust (VV), which is endowed with endoscopic technology, collects the data. TThe VV consists of 4 hospitals and offers medical services to a population of 470.000 individuals. Baerum Hospital, one of the listed medical facilities, has a significant department specializing in gastrointestinal medicine, which served as the primary source for the acquisition of training data.

The KVASIR dataset (Pogorelov et al., 2017) provides a comprehensive compilation of endoscopic operations across the gastrointestinal (GI) tract, encompassing a diverse range of polyps. This dataset is the initial one to faithfully replicate these techniques, with each category including 1,000 images. Eight different types of abnormalities are separated in the dataset: normal-cecum, normal-pylorus, normal-z-line, esophagitis, polyps, ulcerative colitis, dyed-lifted polyps and dyed-resection-margins. Table 5 shows the specific description about the dataset.

The proposed framework is also evaluated using the second dataset (Silva et al., 2014). The dataset has four distinct classifications, namely ulcerative colitis, polyps, normal, and esophagitis. There are 1,000 instances in each class. The examination technique at WCE has been greatly enhanced, leading to improved findings while ensuring patient comfort. The device provides a visual representation of the gastrointestinal system, resembling conventional techniques, and its compact and capsule-shaped design aids medical professionals in identifying issues with more efficiency.

Performance of pretrained deep learning models

The experimental results of fin-tuned pretrained deep learning models are listed in Table 6. In all medical related disease detection, artificial intelligence based deep learning models are attained outclass accuracy as compared to traditional or handcrafted machine features. DenseNet helps solve the problem of vanishing gradients, improves feature transfer, promotes utilizing features, and significantly reduces the total number of para-meters. DenseNet121 used to the Kvasir dataset in order to detect gastrointestinal tracts. The Densenet121 detects the gastrointestinal tracts with an overall and micro-average accuracy of 88% and 89%, respectively. This model achieved a 91% precision and 86% F-measure score for esophagitis. The ulcerative-colitis class achieved 98% highest precision and 92% F-measure. The CNN model has an overall accuracy of 95.22% and a polyps accuracy of 96%. The VGG16 model attained an accuracy of 97.6% and an area under the curve of 99.54%. In comparison to CNN and DenseNet models, VGG16 achieved 98% accuracy on polyps. ResNet50 also achieved great results, while Xception achieved lowest results. Figure 4 presents the performance metrics of pretrained models.

Performance of proposed model using the Kvasir and WCE datasets

The performance of the proposed model using Kvasir and the WCE dataset is presented in Table 7. The proposed model achieved 98.94% accuracy using the Kvasir 8 eight-class dataset and 96.63% accuracy on the WCE dataset. The proposed model obtained the 99.87% maximum AUC score using the Kvasir dataset. All diseases achieved micro-average scores of 99%. The results of the proposed technique is exceptional across all eight classifications. Using kvasir data, only patients with normal-cecum disease obtained a recall rate of 95%, whereas all others achieved a minimum of 97% and a maximum of 100% accuracy. The authors also conducted additional experiments utilizing an additional WCE dataset, and the proposed model obtained an AUC score of 99.58%.

The efficacy of a classification model can be measured using a confusion matrix. If there are unequal numbers of tests in each class is imbalanced or if the data has more than two classes, classification accuracy by itself can be inaccurate and misleading. By creating a confusion matrix and examining the results, authors might be able to better understand the positive and negative aspects of your classification model. The matrix demonstrates the values of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) that were created by the model based on the test or validation data. The confusion matrix is utilized in order to determine the number of true add false predictions for each class. These figures are then organized according to the class that was predicted. In addition, the confusion matrix may be used to generate several metrics, including as recall, precision, and F1 measure, that are employed to analysed the performance of classification models.

Figure 5 illustrates the results of the proposed model using the confusion matrix. In the confusion matrix, horizontal illustrates the actual classes and vertical shows the predicted classes. In the figure, 0 indicates dyed_lifted polyps, 1 dyed_resection margins, 2 esophagitis, 3 normal_cecum, 4 normal_pylorus, 5 normal z_line, 6 polyps, and 7 ulcerative_colitis disease. Class zero makes 99.1% correct predictions, class one makes 100%, class two makes 97.7%, classes three and four make 99.6%, class five makes 99.5%, class 6 makes 98.2%, and class seven makes 100% correct predictions. The most wrong predictions made by the class two mean esophagitis disease, and the most correct predictions were made by the classes one and seven. The proposed model overall makes 960 correct predictions.

Plots that are referred to as learning curves are utilized in order to demonstrate how well a model performs on training set. The training loss metric is used as an indicator of how well a deep learning model fits the training data. To put it another way, it determines how inaccurate the model is based on the data from the training set. Deep model was constructed using only a portion of the dataset, which is referred to as the training set. The amount of training loss can be computed by adding up all of the errors made by each training set sample. Additionally, it is essential to keep in mind that the training loss will be measured at the conclusion of each batch. The training loss curve may often be seen by plotting it, which is one typical way. For the purpose of determining how well a deep learning model operates on a validation set, a metric known as validation loss is utilized. A portion of the dataset known as the validation set has been separated off so that the accuracy of the model can be assesses. The validation loss, which is related to the training loss, is calculated by adding up the number of errors made across all of the samples that are included in the validation set.

These types of curves are essential for evaluating the performance of models on training or validation sets. Figure 6 depicts a variety of learning curve graphs utilizing pretrained models and the proposed model. We have plotted the training and validation curves for five pretrained deep learning models. We observed that the proposed model had the maximum accuracy and lowest loss compared to all other models shown in Fig. 6.

Deep learning models are intrinsically more dependent on processing resources than other techniques. This is due to the fact that deep learning models contain a massive number of parameters and demand a larger amount of training data. It is necessary to calculate how much time will be needed to run the model in the training procedure. Calculating the computational time of deep learning models versus the proposed model is the main emphasis of this study. The computational time of any model is a challenge that must be solved by all researchers working in this field in the modern era. The researchers are able to better determine the relevance of each hyper-parameter they choose for the creation of a model with the assistance of time complexity. Table 8 shows the computational time and loss of the proposed model and the pretrained deep learning models to differentiate how efficient the proposed model is computationally. The computational time of the xception model is 72 s per epoch, and it achieved a 0.62 validation loss, which is the highest among other models. The proposed model achieved a 0.03 validation loss and takes 44 s per epoch.

Figure 7 illustrates the loss caused by the proposed model in comparison to the pretrained deep learning models, hence providing ways to assess the relative efficiency for the current method. The Xception model demonstrates a validation loss of 0.62, surpassing that of other models. The proposed model attained a validation loss of 0.03.

The authors also performed experiments using k-fold cross-validation as presented in Table 9. The authors performed twofold, fivefold, and tenfold validation on the proposed model to check the model’s efficiency. The authors achieved 97.04% accuracy and precision, 97.04% recall, 97.03% F-measure, 99.43% AUC, and 0.138 validation loss with twofold cross-validation. When the authors applied fivefold cross-validation, they achieved 98.78% accuracy and 99.88% AUC with a 0.042 validation loss. The results for tenfold cross-validation show that the proposed model achieved 99.22% accuracy and 0.034 loss. The proposed model performs well even with Kfolds. Training and validation accuracy curves using twofold, fivefold and tenfold validation are shown in Fig. 8A and training and validation loss curves using twofold, fivefold and tenfold cross validation are shown in Fig. 8B.

Table 10 shows the epoch-wise accuracy of the proposed model. The first ten epochs accuracy and the last ten epochs accuracy, as well as their standard deviation (STD), are calculated. The first ten epochs show that epoch 1 has the lowest accuracy and then increases dramatically at each epoch. The highest accuracy is at epoch 7, which is 99.39. The average accuracy of the first ten epochs is between 0.96 and 0.025 STD. The average accuracy of the last ten epochs is between 99.10 and 0.001 STD.

Comparative analysis of proposed model using Kvasir and WCE datasets with the previous studies

Analysis of the proposed model using the Kvasir and WCE datasets in comparison to previous research is shown in Table 11. When employing an 8-class dataset, the CNN was able to obtain an accuracy of 94.83% (Wang et al., 2022). The DenseNet model achieved accuracy values of 95.63%, precision scores of 89.5%, recall scores of 83.5%, and F-measure scores of 82.5% respectively. In order to classify the diseases, they used images obtained via wireless capsule endoscopy (Hosain et al., 2022) . The authors employed 5-class dataset for the detection of gastointesrinal disorders, and they attained an accuracy rate of 96% (Noor et al., 2023). The study Owais et al. (2019) demonstrated a deep learning-based classification-driven retrieval technique. This approach might identify gastrointestinal diseases. A deep learning classification network and identification technique comprise the structure. The classification network analyzes the current medical status to make an informed prediction about the disease, and the retrieval component displays relevant instances from the database. Khan et al. (2022) employed Image-Net weights to fine-tune the MobileNetV2 model for gastrointestinal diseases. Grad CAM and ResNet152 were also employed by the authors to create a comprehensible deep learning system for endoscopic image classification. They used the free Kvasir database’s 8,000 wireless capsule images.

An effective augmentation technique was employed to classify medical images using the heat map of classification results, which had an accuracy of 98.2% during training and 93.46% during validation (Mukhtorov et al., 2023). The previous results on Gastrointestinal tracts demonstrate that the proposed model is outclassed in terms of all performance metrics; it achieved 99.22% accuracy on dataset 1 (eight classes) and 96.63% on dataset 2 (four classes). Also, the proposed model achieved 99.87% and 99.58% AUC, respectively.

Deep learning and optimized EfficientNet models with CNN layers are used in the experimental design to make the distribution of dataset classes more fair. Data augmentation is a practice that artificially increases the size of the training set by creating updated copies of the dataset using existing data. The optimized model, which incorporates the augmentation approach, has proven to be more beneficial in improving model performance. The optimized model achieved the highest level of accuracy, with a quicker processing time of 44 s per epoch compared to the other model. The model’s accuracy ranged from 0.96 to 0.025 standard deviations over the first ten epochs, with the highest accuracy at 99.39 at the seventh epoch. Validation of the model was carried out using the twofold, fivefold, and tenfold methods. The authors achieved 97.04% accuracy and precision with twofolds, 98.78% accuracy with fivefolds, and 99.22% accuracy with tenfold. Despite the presence of Kfolds, the suggested model functions admirably, demonstrating that augmentation strategies are more beneficial for increasing model performance.

RQ1: The choice of techniques for data enhancement significantly improves the performance of an optimized EfficientNet model in the classification of gastrointestinal diseases. Effective use of augmentation techniques like as flipping, rotating, and zooming can increase the model’s ability to generalize and reduce overfitting. As a result, the network can better extract distinctive features from the enhanced data. Therefore, optimizing the augmentation pipeline is critical to finding a balance that boosts the model’s generality while limiting detrimental implications on its capacity to correctly recognize gastrointestinal tracts.

RQ2: The proposed model attained maximum accuracy while requiring few computational resources. The proposed model has a much faster processing time of 44 s per epoch, in contrast to the Xception model, which takes 72 s per epoch. The model’s underlying parameterization enables easy scalability, making it an excellent choice for scenarios with limited resources. The integration of efficiency, cost-effectiveness, and parameter optimization makes it a robust and adaptable solution for classifying gastrointestinal tracts. It meets the demands of real-world applications while maintaining high levels of accuracy and reliability.

RQ3: We validate the efficacy of our proposed model by applying it to a separate dataset for the purpose of classifying gastrointestinal diseases. The results obtained were outstanding, further demonstrating the usefulness of our model. The testing demonstrated its reliability as a versatile method for efficiently and correctly detecting diseases in diverse healthcare environments.