A two-phase transfer learning framework for gastrointestinal diseases classification

Introduction

Gastrointestinal (GI) diseases refer to disorders affecting various parts of the GI tract, which play an essential role in the digestive system. Parts of the GI tract include the mouth, throat, gullet, stomach, bowels, rectum, and anal canal, which form the route for all our nutritional intake (Britannica, 2024). The GI tract functions to digest food, take in nutrients, and remove waste from the body. Abnormalities in the GI tract impact the well-being and healthcare of 40–60% of the global population (Sperber et al., 2021). In developed nations, individuals experience a gastrointestinal illness about once every two years, whereas in developing countries, five to ten cases can occur annually (Pierre Payment, 2002). Additionally, over 50,000 new cases of GI illnesses arise every minute, with some patients, particularly young children, losing their lives consequently. Children between the ages of 1 and 4 years have the highest rate of gastrointestinal diseases.

Internal inspection of the GI tract plays an important role in recognizing various GI diseases, particularly those presenting visible mucosa or tissue damage within the GI tract. Gastroenterologists use various non-invasive and invasive techniques to assess the gastrointestinal tract. Non-invasive approaches include imaging methods like ultrasound, CT scans, magnetic resonance imaging (MRI), and X-rays (Frøkjær, Drewes & Gregersen, 2009), along with other functional assessments such as gastric emptying studies, esophageal manometry, pH monitoring, and breath analysis tests (Schwizer, Fox & Steingötter, 2003). Endoscopy is the most common invasive method to examine the GI tract, involving the insertion of light and a camera into the GI tract using a flexible tube through a natural entrance like the mouth. Various endoscopic techniques are used depending on the specific section of the GI tract to be inspected. For instance, esophagogastroduodenoscopy (EGD) examines the upper GI tract, including the esophagus, stomach, and duodenum. Similarly, colonoscopy is used to view the entire large intestine (colon), and sigmoidoscopy is used for inspecting the lower colon and rectum (Banks, 2017). Additional invasive techniques include endoscopic ultrasound (EUS), which combines endoscopy with ultrasound (Fabbri et al., 2014). Capsule endoscopy involves swallowable capsules with cameras to capture images as they travel through the GI tract (Ciuti, Menciassi & Dario, 2011). Endoscopic retrograde cholangiopancreatography (ERCP) merges endoscopy with X-ray imaging to help diagnose and treat various abnormalities (Jorgensen et al., 2016).

Recent developments in computer vision upgrade healthcare diagnostic systems, which assist in automatically detecting various medical conditions, including GI diseases. A key use of computer vision is applying convolutional neural networks (CNNs) to analyze medical images like computed tomography (CT) scans, endoscopic images, magnetic resonance imaging (MRIs), and X-rays. This analysis of medical images helps healthcare specialists in making early diagnoses. Early contributions to computer-aided diagnosis of GI diseases relied on standard approaches, such as feed-forward neural networks to predict conditions like lymph node metastasis in stomach cancer (Hensler et al., 2005). However, with advancements in deep learning methods, GI diagnostic systems are enhanced by integrating CNN-based architectures, such as InceptionNet and VGGNet (Haile et al., 2022). In addition, more recent approaches use attention mechanisms and hybrid methods to address the complexities of large-scale endoscopic data analysis (Lonseko et al., 2021; Fati, Senan & Azar, 2022). Ensemble learning approaches, such as the GastroVRG model proposed by Islam, Rony & Sultan (2024), combine multiple classifiers like VGG, random forest, and gradient boosting to achieve high performance in identifying conditions like esophagitis and polyps.

The proposed study aims to develop a transfer learning framework to train high-performing image classification models for the timely and accurate detection of GI diseases using endoscopic images. This study’s motivation comes from the role of endoscopy in diagnosing GI diseases. However, manually analyzing endoscopic images is a challenging and time-consuming task. It requires extended knowledge and experience in the field of gastroenterology. Additionally, the increasing use of endoscopic procedures and the growing number of images produced add more pressure on healthcare systems. Moreover, a study predicted a shortage of gastroenterology experts by 2020, creating congestion in the diagnostic process and delaying diagnoses for patients in need (Sullivan, 2009). Given these challenges, the motivation is to contribute to developing automated endoscopy systems that could assist gastroenterologists in analyzing endoscopic images more efficiently. Deep CNN models can learn complex visual patterns from image datasets. When integrated with endoscopy systems, these models can automatically identify GI diseases in endoscopic images and reduce the time required for diagnosis. Thus, easing the diagnostic burden on gastroenterologists and ultimately helping them make more informed and timely clinical decisions. Our primary contributions are as follows:

• A two-phase transfer learning framework is proposed that utilizes pre-trained CNN models and specifically fine-tunes them for clinical use in gastroenterology. Xception, InceptionResNetV2, and VGG16 are employed to classify gastrointestinal diseases using endoscopic images.

• The proposed method demonstrates the advantages of fine-tuning pre-trained models, significantly improving consistency and performance. The quantitative results demonstrate notable performance improvements compared to fully trained models.

• The proposed framework is flexible and adaptable to various classification tasks in gastroenterology. It can be modified easily for different clinical scenarios by organizing training images into appropriate directories for each new task.

Related work

Many research studies have been conducted on diagnosing gastrointestinal (GI) diseases through machine learning and deep learning. Table 1 provides a summarized overview of related research work. Although many studies have shown encouraging outcomes, there is still a need to enhance the generalization of AI models, which can be achieved by increased diversity in the training data (Niu et al., 2020). Endoscopic data from different gastroenterology departments worldwide should be gathered, and expert gastroenterologists should label the data.

Early efforts in computer-aided diagnosis for detecting and classifying GI tract diseases set the basis for recent developments. For example, Hensler et al. (2005) developed a neural network for the pre-operative identification of lymph node metastasis (LNM), a factor used to predict prognosis in stomach cancer. The approach performed better than earlier methods, achieving an accuracy of 72.73%. Building on these foundations, recent approaches use CNNs to classify GI tract diseases. Hirasawa et al. (2018) created a CNN-based diagnostic system trained on over thirteen thousand endoscopic images of GI diseases. A separate dataset of approximately 2,300 endoscopic images is used to test the trained model. The model took less than 50 s to analyze all the images, achieving an overall true-negative rate of 0.922 for polyps in stomach cancer. The study concluded that the proposed model is clinically relevant to assist gastroenterologists in their daily practice. Haile et al. (2022) proposed a technique that combines features extracted from InceptionNet and VGGNet convolutional networks. These combined features are classified using various techniques, including softmax function, random forest, KNN, and SVM. Results show that SVM achieved the highest classification accuracy of 98%.

Further advancements in this field focus on analyzing large volumes of endoscopic video data. Fati, Senan & Azar (2022) worked on the early detection of lower GI tract tumors using endoscopic videos. The study highlights the challenges in analyzing a large number of frames produced in endoscopic videos and suggests using artificial intelligence to help make accurate diagnoses. Four systems utilizing various approaches, including feed-forward and artificial neural networks, pre-trained models like AlexNet and GoogLeNet, hybrid techniques combining CNN models with SVM, and hybrid feature extraction methods incorporating traditional algorithms such as the gray-level co-occurrence matrix, local binary pattern, and fuzzy color histogram are presented. The proposed solution shows better performance in diagnosing tumors in endoscopic video frames. The feed-forward neural network utilizing hybrid features achieves an outstanding accuracy score of 0.993 and an AUC of 0.998.

Attention mechanisms in CNN architectures also show promising results. Lonseko et al. (2021) proposed a deep CNN utilizing attention mechanism for GI disease classification, implemented with encoder and decoder layers. This mechanism performs better than earlier classification methods, which often miss complex spatial dependencies in images due to their structural limitations. The attention model achieves an accuracy of 93.1% and F1-score of 0.928.

Several studies have also explored other feature extraction and optimization methods for classifying GI diseases. For example, Azar et al. (2023) focuses on classifying large intestine cancer, which is among the deadliest cancers. The study examines six deep-learning optimizers, with empirical results revealing CNN-Adam as the overall top performer, achieving 82% average accuracy across four large intestine cancer datasets. In one particular dataset (Dataset-1), CNN-Adam achieves an accuracy of 95%, compared to 76% with RMSprop and 96% with Adadelta. Ali et al. (2018) introduced a feature extraction method combining a Gabor-based gray-level co-occurrence matrix (G2LCM) with an SVM classifier to classify GI diseases in chromoendoscopic images. Their model classifies three classes of GI chromoendoscopic images-one normal and two abnormal- and achieves an accuracy of 91%, a sensitivity of 0.91, a specificity of 0.82, and an AUC score of 0.87.

Transfer learning has also shown effective results in enhancing the performance of CNN models for their use in medical image classification. Bang, Cho & Baik (2020) uses a pre-trained InceptionResNetV2 model and fine-tunes it for a custom endoscopic dataset with five classes of GI tract. The model achieves a weighted accuracy score of 0.846. In the prospective validation phase, the model achieves a lower score of 0.764 than the expert endoscopist with a score of 0.876. Similarly, Yogapriya et al. (2021) used pre-trained CNN models and fine-tuned them to classify GI tract diseases. GoogLeNet, VGG16, and ResNet18 are used as base pre-trained models. The VGG16 achieves the best performance with an accuracy score of 96.33%, effectively differentiating among eight classes of endoscopic images from the GI tract.

Ensemble learning approaches have also been used to improve classification performance. Islam, Rony & Sultan (2024) proposed a novel approach called GastroVRG, which helps in the early screening of GI diseases using advanced transfer features. VGG16, random forest, and gradient boosting models are used to extract features for classifying esophagitis and polyps. The gradient boosting model achieves a classification accuracy of 99.73% when tested on the Kvasir-v1 dataset.

The wireless capsule endoscopy (WCE) is complex and time-consuming because it requires manual review of a large number of frames produced. Researchers have worked on developing automated diagnostic systems for recognizing GI abnormalities through WCE. Majid et al. (2020) proposed a system that includes five key steps: creating a database, extracting handcrafted and CNN deep features, combining features, feature selection by a genetic algorithm, and recognizing diseases. The extracted features include cosine, wavelet, color, and VGG16 feature maps, combined and optimized through a genetic algorithm. An ensemble classifier is then utilized for disease recognition. This technique outperformed existing methods, achieving an accuracy of 96.5% for four types of GI abnormalities, including ulcers, polyps, esophagitis, and bleeding. More recently, Malik et al. (2024) developed four multi-classification deep learning models. The models include VGG19 paired with custom CNN layers, ResNet152V2, gated recurrent unit (GRU) paired with ResNet152V2, and ResNet152V2 paired with bidirectional GRU (Bi-GRU). The models are trained to identify ulcerative colitis, polyps, and dyed-lifted polyps from WCE images. VGG19 paired with custom CNN layers achieved the highest accuracy of 99.45%.

Recent studies have also explored the use of advanced deep learning architectures. Mulugeta Abuhayi et al. (2024) introduced a method that uses recurrent vision Transformers (RVT) combined with wavelet transform to classify GI diseases from endoscopic videos. Their approach includes three key phases: processing images, extracting features, and classification. Wavelet transform is used for segmentation, and adaptive median filtering reduces noise in the image processing phase. Feature extraction is performed using concatenated RVT, and classification is carried out using an ensemble of classifiers, including SVM, Bayesian networks, random forest, and logistic regression. The Hyper-Kvasir dataset is used to test the model. The model achieves an accuracy of 99.13% and an AUC of 0.9954. Another study conducted by Asif et al. (2024) introduced the DCDS-Net model, a deep transfer network that integrates depth-wise separable convolution with densely connected blocks (DCB) and residual connections. The model addresses issues such as overfitting and enhances generalization by integrating Dropout, Batch Normalization, Global Average Pooling, and Dense layers within the DCB. A dataset of 6000 endoscopic images across four classes of GI diseases is used to evaluate the model. DCDS-Net shows high performance, achieving a classification accuracy of 99.33%, a precision of 0.9937, and a recall of 0.9932.

Proposed method

Data acquisition and description

Kvasir-v2 (Pogorelov et al., 2017) is the primary dataset for the proposed research. This dataset was collected from a large gastroenterology center in Norway. Expert endoscopists have labeled and verified the dataset to ensure high-quality annotations. We obtained the dataset from a publicly available repository known as Simula Datasets. There are 8,000 images in the dataset, equally divided into eight categories of the gastrointestinal tract. These include three categories of anatomical landmarks, three categories of pathological findings, and two categories related to polyp removal, described below.

Anatomical landmarks

Anatomical landmarks are visual features in the GI tract that can identify the location of any particular finding for navigation. The three classes of anatomical landmarks are Z-line, pylorus, and cecum. (A) The z-line is a visual boundary between the esophagus and the stomach. During an endoscopy, it appears as a sharp transition from the esophagus’s light mucosa to the stomach’s darker mucosa. (B) The pylorus refers to the region surrounding the opening at the end of the stomach where the small intestine starts. (C) The cecum is pouch-like, and the starting section of the large intestine is located immediately after the small intestine. Figure 1 presents an endoscopic sample for each class in the anatomical landmarks category.

Pathological findings

Pathological findings are visually affected sections in the GI tract. During an endoscopy, these can be observed as abnormal features from the normal mucosa’s texture, size, and color. These abnormalities either indicate any current disease or the starting of a new disease, possibly leading to cancer. Early detection of pathological findings in a GI patient is crucial for starting treatment timely. The three classes of pathological findings are esophagitis, polyps, and ulcerative colitis. (A) Esophagitis is a noticeable swelling in the esophagus, the tube connecting the mouth to the stomach. This inflammation can be caused by various reasons, such as side-effects of oral medication, allergic reactions, viral/bacterial infections, and conditions like gastroesophageal reflux disease where stomach acid flows backward in the esophagus (Grossi, Ciccaglione & Marzio, 2017). (B) Polyps are growths inside the GI tract that develop into abnormal tissues (Waldum & Fossmark, 2021). Polyps appear in various sizes and shapes; some may indicate early signs of cancer, while most do not develop into cancer (Goddard et al., 2010). (C) Ulcerative colitis is a long-term inflammatory disease in the intestines that leads to swelling and sores in the colon and rectum (Langan et al., 2007). Figure 2 presents an endoscopic sample for each class in the pathological findings category.

Polyp removal

Two classes in the polyp removal category can assist in accurately identifying dyed or lifted polyps for removal and post-removal analysis. (A) Dyed and lifted polyps help identify the location of dyed or lifted polyps, making it easier to separate them from normal tissue during removal. (B) Dyed resection margins are used to determine the post-removal status of the polyps. Figure 3 presents an endoscopic sample for both classes in the polyp category.

Esophagitis severity

In addition to the Kvasir-v2 dataset, a subset of the Hyper-Kvasir dataset (Borgli et al., 2020) is also used. This is to explore a potential clinical application for binary classification of different severity grades of esophagitis disease. The Hyper-Kvasir dataset contains a comprehensive classification of annotated endoscopic images of the GI tract, broadly categorized into upper and lower GI tract. In the upper GI tract, two classes of images are related to the severity of esophagitis. These classes are esophagitis A and esophagitis B-D, as shown in Figs. 4A and 4B. Esophagitis A refers to lower severity of esophagitis, while esophagitis B-D refers to higher severity esophagitis. This system for grading severity is known as the Los Angeles classification system (Lundell et al., 1999), where mucosal breaks under 5 mm are classified as esophagitis A, and those over 5mm fall under esophagitis B, C, or D, with D being the most severe. Automated binary classification between esophagitis A and esophagitis B–D can help differentiate between lower and higher severity levels, supporting early treatment for grade A esophagitis. This data is also obtained from the same repository (Simula Datasets; https://datasets.simula.no/hyper-kvasir/).

Data preprocessing

This study utilizes the computational resources provided by Google Colaboratory (Bisong, 2019), which offers a cloud-based environment for running Python code, including machine learning tasks. The dataset, available in structured directories, is directly downloaded into the Colab environment for processing. It is first divided into two main parts: 10% is randomly selected and set aside for testing the final trained models, while the remaining 90% is used for training and validation. From this 90%, 70% is allocated for training the model, and the remaining 20% is used for validation, allowing us to monitor and adjust the model’s performance during training. The initial 10% is reserved exclusively for testing the final trained models, ensuring an accurate evaluation of their ability to generalize to new, unseen data.

Three-channel RGB images in the dataset have different dimensions, ranging from (720, 576, 3) to (1,920, 1,072, 3). To ensure uniformity during training, the images are resized to maintain a 1:1 aspect ratio, in line with the input requirements of different model architectures. For example, input images are resized to (299, 299, 3) for the InceptionResNetV2 model and (224, 224, 3) for the VGG16 model. This resizing process ensures that the images are compatible with the different CNN architectures while preserving essential visual information. Figure 5 provides a detailed overview of the end-to-end workflow, highlighting each stage of the process from dataset acquisition to model training, testing, and performance evaluation to facilitate understanding of the entire pipeline. This workflow is followed in all the classification tasks for both the Kvasir and the Hyper-Kvasir datasets.

Transfer learning framework

Transfer learning is a common method in deep learning, particularly in cases where the available dataset is small or specialized. It uses pre-trained models originally optimized for general classification tasks. These models are generally pre-trained on popular datasets, such as ImageNet (Deng et al., 2009) or Microsoft COCO (Lin et al., 2014). Transfer learning adapts the pre-trained models to solve a different classification task. One of the main benefits of transfer learning is that the pre-trained models have already learned important features, and they can be repurposed for a new task, reducing the need for extensive computational resources and data. Several models pre-trained on ImageNet data are employed as the base models for the proposed method. These base models serve as feature extractors for the new task of classifying GI tract endoscopic images. The architecture of these models is customized by removing original classification layers and adding task-specific layers on top of the pre-trained base model. These additional layers are added to meet the specific requirements of the classification task, particularly modifying the final dense layer to match the number of classes and adjusting the activation function based on the environment setup. For example, a softmax activation function is applied in the output layer for multi-class classification. The proposed framework uses a two-phase training process outlined in Algorithm 1.

Algorithm 1:

Workflow with the proposed two-phase transfer learning process.

Input: Endoscopic Images

Output: Trained Models and Evaluations

Data Partitioning

$testingData\leftarrow$ 10% data

$trainingData\leftarrow$ 70% data

$validationData\leftarrow$ 20% data

Load Pre-trained Base Model

$numClasses\leftarrow$ len(trainingData.classNames)   ▹ number of output classes in a classification task

$baseModel\leftarrow$ (keras.applications.PreTrainedCNN:

weights = "imagenet"

include_top = False)   ▹ pre-trained model with ImageNet weights using Keras applications module. Original classification layer is not included.

$baseModel.trainable\leftarrow False$   ▹ Freeze the layers in the base model

Add Custom Layers

$x\leftarrow baseModel(inputs)$

$x\leftarrow$ GlobalAveragePooling2D(x)

$x\leftarrow$ Dropout(0.2)(x) ▹ Add dropout layer to prevent overfitting. 0.2 means that 20% of the nodes will be randomly set to zero during each training epoch.

$outputs\leftarrow$ Dense(numClasses, softmax activation)   ▹ Output dense layer automatically gets the number of nodes equal to the number of output classes in the respective classification task

Compile the Model

$model\leftarrow inputs$, outputs

$optimizer\leftarrow$ Adam

$lossFunction\leftarrow$ Sparse Categorical Entropy

$learningRate\leftarrow$ 0.001

$evaluationMetric\leftarrow$ Accuracy

$model.compile\leftarrow$ (optimizer, lossFunction, learningRate, evaluationMetric)

Train Model Custom Layers (Phase 1)

$epochs\leftarrow 50$

$model.fit\leftarrow$ (trainImages, validImages, epochs)

Fine-tune the Entire Model (Phase 2)

$baseModel.trainable\leftarrow True$   ▹ Unfreeze base model layers

$learningRate\leftarrow$ 0.00001

$model.compile\leftarrow$ (optimizer, lossFunction, learningRate, evaluationMetric )   ▹ Recompile the model with a very small learning rate

$epochs\leftarrow 20$

$model.fit\leftarrow$ (trainImages, validImages, epochs)

Test the Trained Model

$predictions\leftarrow$ model.predict(testImages)   ▹ Test predictions are used to evaluate the model

DOI: 10.7717/peerj-cs.2587/table-13

Figure 6 demonstrates the architecture of the proposed two-phase transfer learning framework, illustrating how the pre-trained models are integrated and customized for classifying endoscopic images of the GI tract.

• 1)

  In the initial training phase, a technique known as “freezing” the pre-trained layers is applied. This means that during backpropagation, only the newly added custom layers are updated, while the weights of the base model remain unchanged. By freezing the pre-trained layers, it is ensured that the model focuses on learning task-specific features from the endoscopic images of the GI tract while still benefiting from the general features (e.g., edges, textures, shapes) learned from the large-scale ImageNet dataset. The custom layers are trained with a higher learning rate to accelerate the convergence of the model’s parameters to the optimal values for the new task.

• 2)

  After sufficient training of the custom layers, models are fine-tuned in the second phase. In this phase, the entire model, including the pre-trained layers, is unfrozen and trained on the endoscopic image dataset. Fine-tuning is performed with a very small learning rate (0.00001 for the proposed setup) to ensure that the weights of the pre-trained layers are adjusted carefully, preventing drastic changes that could lead to overfitting or loss of the general features learned from the ImageNet dataset. This phase aims to refine the entire model, allowing the pre-trained layers to adapt to the specific characteristics of the GI tract images while retaining the benefits of transfer learning. This process enhances the model’s ability to generalize to unseen data.

Results

Discussion

For the first classification task, gastrointestinal tract classification, Table 7 results indicate that all five models correctly classify more than 90% of the endoscopic images in the training subset. Classifying endoscopic images in the validation subset, the three fine-tuned models reduce their performance by 5–10%, which is a naturally expected behavior. However, the custom CNN models significantly reduce their performance by more than 20%. This difference between models’ performance on the training images and validation/testing images helps evaluate the generalization behavior of the models. This behavior is known as overfitting when the models show excellent classification performance on the images they were trained on and significantly reduce their performance when tested for unseen images. This behavior is undesirable since it can lead to models’ poor generalization towards unseen data. Figure 7A presents a detailed analysis of models’ training and validation performance for the first classification task throughout the training process. The BasicConvNet exhibits the highest level of overfitting, likely because of its simpler architecture. By adding a few batch normalization and dropout layers to the BasicConvNet architecture, the RegConvNet shows a slight improvement in maintaining higher validation accuracy and reducing overfitting due to these regularization methods. However, it still falls short of ideal performance, particularly for a healthcare classification task with eight categories. The fine-tuned Xception, VGG16, and InceptionResNetV2 models show very similar accuracy scores during training and perform significantly better on unseen images. Another important observation is that the custom CNN models show highly fluctuating performance during training, providing the same learning rate parameter. This is generally caused by insufficient training, and since the custom CNN models are fully trained from scratch, they do not completely learn the underlying patterns from a limited number of endoscopic images. Contrarily, the pre-trained models achieve more consistent performance throughout the training. This benefit is due to their pre-trained parameters and more optimized architecture. The fine-tuning process, depicted in Fig. 7B, further enhances the performance of only pre-trained models. The three fine-tuned models achieve comparable accuracy and generalization performance, with the fine-tuned InceptionResNetV2 showing the best results and the slightest difference between its training and validation accuracies.

For the second classification task, pathological findings classification, Table 8 results indicate that all five models correctly classify more than 95% of the endoscopic images in the training subset. This classification accuracy slightly decreases on the validation and testing subsets by 1–3% for the three fine-tuned models and 10–15% for the custom CNN models. Due to the small number of classes, all five models perform significantly better than the previous classification task. However, the trends and patterns during the training process are quite similar to the previous task, as presented in Fig. 8A. The same overfitting behavior is observed for the BasicConvNet and RegConvNet, with RegConvNet showing even more performance fluctuations. These may also be due to the nonoptimal integration of regularization techniques such as dropout and batch-normalization. The BasicConvNet shows the highest level of overfitting, which shows its inability to generalize well for a three-class classification problem due to an even smaller training dataset. The results also indicate that tuning the hyperparameters of deep learning models can differ across different use cases, even within the same problem area. The transfer learning models demonstrate much higher and comparable performance regarding generalization, achieving an increase of over 10% in accuracy scores compared to the earlier classification task. The InceptionResNetV2 again persists with the highest performance, presenting the same behavior during the fine-tuning process, as presented in Fig. 8B. These results demonstrate that properly fine-tuned transfer learning models are capable and reliable enough to be integrated into modern endoscopy systems for such classification tasks.

Moving on to discussing polyp identification results, all five models show excellent accuracy scores for this less challenging binary classification task. Table 9 shows that all five models achieved desirable accuracy scores during training/fine-tuning, validation, and testing. However, looking at Fig. 9A for a detailed analysis, performance fluctuations are extremely elaborated for the RegConvNet. RegConvNet exponentially learns to classify polyps and normal pylorus images during the five training epochs. However, it fluctuates, especially around epochs 25–30, significantly dropping performance. RegConvNet achieves maximum validation and testing accuracy scores of 98.2% and 95.5%, respectively, which are not bad. However, relying upon such models for real-world clinical situations is not recommended, and proper model convergence should be considered as the primary deciding factor. One surprising observation is that the BasicConvNet, the simplest model, achieves better testing accuracy than RegConvNet and even the fine-tuned VGG16. In terms of maintaining performance consistency throughout the fine-tuning phase, Xception shows higher validation scores and slightly excels over the fine-tuned InceptionResNetV2 and fine-tuned VGG16, as presented in Fig. 9B. The fine-tuned InceptionResNetV2 achieves 100% training accuracy and excellent 99.5% testing accuracy, even more than its validation accuracy of 98.9%. This means that out of 200 testing images of polyps and normal pylorus, the fine-tuned InceptionResNetV2 correctly classifies 198 images. This demonstrates the high reliability of a properly fine-tuned InceptionResNetV2 in such clinical scenarios.

For the esophagitis severity classification, results presented in Table 10 indicate overfitting behavior by all five models. Most models, except BasicConvNet, show good training accuracies but significantly reduce their performance on validation and testing images. Contrary to their performance in the previous three classification tasks, all three fine-tuned models also drop their performance by more than 15% during testing. As presented in Fig. 10A, all five models show the most inconsistent performance and achieve the least validation accuracies. Esophagitis severity classification represents a binary classification problem, and the models are expected to perform better. However, they do not. This unexpected behavior appears to be due to the high resemblance between the images of esophagitis A and esophagitis B–D. Let us closely examine the endoscopic samples presented in Fig. 4. We can see that they are quite similar in color and texture features because they come from the same organ, the esophagus. Regarding the fine-tuning process presented in Fig. 10B, all three transfer learning models present better consistency due to a smaller learning rate parameter. The fine-tuned VGG16 slightly excels over the fine-tuned InceptionResNetV2 and fine-tuned Xception. Considering the best performing fine-tuned InceptionResNetV2 with a testing accuracy of 74.2% means that out of 66 unseen images, it correctly classifies around 49 images. This is way less than the desired performance in clinical scenarios, where high misclassification rates lead to high risks in diagnoses and treatments.

Moreover, the models are evaluated based on other weighted average performance metrics to ensure that primary reliance on accuracy metrics is significant. Table 11 compares weighted average metrics for five models across four classification tasks.

It is important to mention that these metrics are obtained from testing the models on unseen images. The first noticeable trend is that the transfer learning models perform better than custom CNN models in most classification tasks. In case of the general classification of endoscopic images of the GI tract in the first task, the fine-tuned VGG16 achieves the highest $W{A}_{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}$, $W{A}_{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$, and $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ of 0.88, followed by 0.86 of both fine-tuned Xception and fine-tuned InceptionResNetV2. Similar to the accuracy comparison, the custom CNN models are behind the fine-tuned models. In the case of pathological findings classification, the fine-tuned InceptionResNetV2 leads with its excellent weighted average score of 0.98, followed by fine-tuned Xception and fine-tuned VGG16. Here, both custom CNN models achieve same reasonable $W{A}_{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}$, $W{A}_{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$, and $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ of 0.87. As for the polyp identification, all the models achieve more than 0.95 for all metrics, with the fine-tuned InceptionResNetV2 again leading the chart with $W{A}_{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}$ of 1.0, $W{A}_{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$ and $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ of 0.9. Similar to lower accuracy scores, all the models achieve lower weighted average scores of below 0.75 in the case of esophagitis severity classification. The fine-tuned InceptionResNetV2 again achieves the highest scores, $W{A}_{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}$ of 0.72, $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ and $W{A}_{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$ of 0.73. The fined-tuned VGG16 shows lower accuracy as compared to the other two fine-tuned models, and performance trends are the same with $W{A}_{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}$ and $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ of 0.66, and $W{A}_{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$ of 0.67. These weighted metrics are useful in evaluating models in image classification problems, especially when a class imbalance exists. These metrics are also in line with the accuracy scores, which demonstrate the significance of these results.

In terms of contributing towards automating the diagnoses of GI diseases based on endoscopic images, the proposed study is generally similar to studies cited in the related works table. However, in terms of methods and data, there are certain differences. Some studies use different datasets for GI tract image classification, while others vary in using diverse AI methods. To make a suitable comparison with previous studies, Bang, Cho & Baik (2020) and Yogapriya et al. (2021) are selected based on their relevance for using transfer learning methodology and fine-tuning pre-trained models. Table 12 compares previous studies and two proposed methods. For general classification of endoscopic images of the GI tract, the fine-tuned VGG-16 model by Yogapriya et al. (2021) achieves benchmark performance with an accuracy of 96.33%. A data augmentation strategy is also utilized to expand the number of images in the kvasir-v2 dataset. This increase in data size helps improve the general classification performance of VVG16. This also highlights the effectiveness of using pre-trained CNNs and fine-tuning them for specific medical imaging tasks. Bang, Cho & Baik (2020) fine-tuned the InceptionResNetV2 for five-category classification achieving a weighted accuracy of 84.6%. For specific disease classification, such as classifying a few diseases or identifying polyps from normal pylorus, the proposed fine-tuned InceptionResNetV2 model shows superior performance. The model achieves an accuracy of 97.6% and $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ of 0.98 for classifying three diseases. For polyp identification, the model achieves an accuracy of 99.5% and $W{A}_{\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}$ of 0.99, showing the highest and most consistent performance in examined classification tasks. The results suggest that the separate fine-tuning of pre-trained models for specific clinical use cases improves the classification performance. This approach allows the proposed models to capture the intra-variabilities and inter-variabilities of endoscopic images, thus achieving higher performance. The complex architecture of the InceptionResNetV2 model, which integrates both inception modules and residual connections, contributes to its superior performance. The inception modules allow the network to capture multi-scale features, while the residual connections help solve the vanishing gradient problem. The ability of InceptionResNetV2 to consistently achieve higher classification accuracies compared to other models can be attributed to its robust feature extraction capabilities and efficient training dynamics. More importantly, the proposed fine-tuned InceptionResNetV2 model excels in more disease-specific classification tasks. This demonstrates the potential of advanced CNN architectures, when appropriately fine-tuned, to enhance diagnostic accuracy in medical imaging, thereby supporting more reliable and effective clinical decision-making.

Conclusion and future work

Gastrointestinal diseases affect a significant population worldwide. Endoscopy is a common method used by gastroenterologists to examine GI diseases internally. However, manual analysis of a large number of endoscopy images is challenging and time intensive. A two-phase transfer learning framework is proposed to fine-tune pre-trained CNN models to classify endoscopic images of the GI tract. Xception, InceptionResNetV2, and VGG16 have been fine-tuned on Kvasir and Hyper-Kvasir datasets. Two custom CNN models have also been trained for comparative analysis. These models have been tested comprehensively across four classification tasks to explore various clinical scenarios where computer vision can assist in endoscopy. The results show that the transfer learning models perform more consistently in most classification tasks. Transfer learning models also generalize much better due to pre-trained weights, even with small datasets, which is usually a challenge in healthcare imaging. Results also show that these models perform better on specific disease classification tasks than on general classification tasks. Moreover, it is essential to properly plan and optimize regularization techniques, such as dropout and batch normalization, when needed. Improper use of such techniques can negatively impact the model’s performance by increasing the fluctuations and decreasing the convergence during training, leading to an overfit model. In conclusion, the results show that properly fine-tuning pre-trained CNN models can effectively diagnose gastrointestinal tract abnormalities reliably. According to the results, InceptionResNetV2 achieved the highest and most consistent performance across different classification tasks. This shows its potential for use in real-world clinical settings, helping gastroenterologists diagnose and start treatment early to lower mortality rates from gastric diseases. However, there are certain limitations where the clinical scenario of classifying the severity grade of the same disease, like esophagitis, needs to be further explored. For example, the attention and segmentation mechanisms can focus on the size of the infected area since other features, like color, shape, and texture, are mostly the same and not helpful in differentiating lower severity from higher. Furthermore, comprehensive and diverse datasets with high-quality annotated images representing a wide range of GI tract features of different patients would enable further enhanced GI disease detection performance.

Supplemental Information