Endoscopic Image Processing: Evaluating the Robustness of Deep Neural Networks

The Hybrid CNN-Transformer Model (HCNNT) is a novel model which exploits Convolutional Neural Networks (CNN) for spatial feature extraction and Transformers for capturing contextual and long range dependencies in order to carry complex image classification. In order to evaluate the algorithm proposed, a new model has been introduced and trained on a data set of gastrointestinal (GI) image classification, which is derived from a balanced data set with 226 images available on Kaggle. Although the cascaded CNN part is to learn elaborate visual characteristics, the improved understanding of complex patterns and relationships is achieved by the Transformer component to process these features. When both are integrated, we gain a powerful system that has high accuracy in distinguishing GI bleeding condition from normal condition. Extensive training and optimization of the hybrid model using binary cross-entropy loss and Adam optimizer has achieved state-of-the-art performance. Through the experiments evaluated on the HCNNT model, it is evident that an accuracy of 97.34% proves to be significantly beneficial, its practicality in terms of reliability for medical image classification that essentially improves the accuracy of diagnosis in gastrointestinal healthcare.