Advancing Colonoscopy Diagnosis for Polyp Recognition Using Fastai and EfficientNet Deep Learning

Introduction Colorectal cancer is one of the global health threats and ranks among the deadliest diseases worldwide. The recognition and elimination of polyps at a primary, precancerous phase are, therefore, the key to preventing CRC. Most of these polyps differ in size and level of malignancy, thus failing to be detected by the commonly used screening methods. Methods In this study, AI-driven tools were designed using deep learning models, such as VGG16, ResNet, and EfficientNet. They were validated using datasets obtained from JSS Hospital to enhance the accuracy of polyp recognition and decrease the probability of CRC, thereby improving patient outcomes. Fastai comes with an intuitive API, where most functions related to data preprocessing, building, and training a model are already built. Logs on training and validation losses, accuracies, and confidence scores of the performance metrics ensure the rigors of evaluation across multiple epochs of training. Results The results were impressive, with the deep learning models performing almost constantly at an accuracy of 99% in image classification. The robustness of the models is guaranteed because the balance between validation loss and training loss is attained. Hence, there is no overfitting or underfitting, guaranteeing reliable predictions. An interactive web platform was developed using Hugging Face with Gradio, and real-time predictions could be made by allowing users to upload images. Discussion The confusion matrix indicated that these models achieved nearly perfect classification performance. The VGG16 model performed with 99.48% accuracy, 100% precision, 97.95% recall, and an F1 score of 98.96%. The VGG19 model outperformed the former by a slight margin, displaying an accuracy of 99.69%, precision of 100%, recall of 98.76%, and an F1 score of 98.37%. ResNet18 and ResNet50 performed exceptionally well, achieving 99.79% accuracy, 100% precision, 99.17% recall, and 99.58% F1 score. The model with the best performance, with a solid score, was Efficient Net, scoring an accuracy of 99.9%. Conclusion In the study, the effectiveness of deep CNN models was validated for polyp detection to aid in CRC prevention. These effectively and stably well-performing models, being provided to users by a very user-friendly platform, set a very good precedent for their broad application in the future. This milestone success and careful evaluation have led to an improvement in diagnostic processes and, therefore, health outcomes.