GIT-Net: An Ensemble Deep Learning-Based GI Tract Classification of Endoscopic Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate classification of gastrointestinal (GI) diseases from endoscopic images is crucial for effective treatment.
• Individual pre-trained models may struggle to capture the full spectrum of disease characteristics, leading to misclassifications.

Purpose of the Study:

• To develop and evaluate an ensemble model for improved classification accuracy of GI endoscopic images.
• To propose GIT-NET, a weighted average ensemble model, for classifying eight types of GI tract diseases.

Main Methods:

• An ensemble model, GIT-NET, was developed by combining predictions from pre-trained DenseNet201, InceptionV3, and ResNet50 models.
• Two ensemble techniques were explored: model averaging and weighted averaging.
• The models were evaluated on the KVASIR v2 dataset.

Main Results:

• Individual models achieved accuracies of 94.54% (DenseNet201), 88.38% (InceptionV3), and 90.58% (ResNet50).
• The model averaging ensemble achieved 92.96% accuracy.
• The weighted average ensemble (GIT-NET) achieved a superior accuracy of 95.00%.

Conclusions:

• The weighted average ensemble approach (GIT-NET) effectively overcomes the limitations of individual models in classifying GI endoscopic images.
• Ensemble methods significantly enhance the accuracy of GI disease classification, leading to more reliable diagnoses.