Beyond X-ray: Deep Learning Solutions for Gastrointestinal Bleeding

Gastrointestinal (GI) bleeding is a critical medical condition that has a substantial impact and poses significant challenges for the global healthcare system due to its varying severity. GI bleeding can occur ranging from mild, self-limiting cases to severe, poten- tially fatal events, which emphasizes the need for timely and accurate diagnosis. Traditional feature extraction techniques, such as color thresholding and Scale-Invariant Feature Transform (SIFT), combined with classifiers like Support Vector Machine (SVM), often fail to detect subtle patterns within complex gastrointestinal images, especially those images captured from wireless capsule endoscopy (WCE).This work demonstrates the distinct benefits of deep learning techniques over conventional methods, especially their capacity to identify minute bleeding patterns in intricate gastrointestinal pictures. Better performance is shown by models like EfficientNetB0, which are computationally efficient, and the custom CNN, which is designed for binary classification. These discoveries aid in the creation of automated, trustworthy diagnostic instruments for use in real-time healthcare settings. Especially, In this research, we introduce the most sought-after deep learning methods namely Custom Convolutional Neural Network (CNN), EfficientNetB0, ResNet50, and InceptionV3. These models and their working have been thoroughly understood and evaluated across various parameters such as precision, accuracy, recall F1-score, etc., These algorithms have proven capability in identifying patterns that traditional methods tend to miss out. The results prove the potential of the models, particularly EfficientNetB0 and custom CNN, which offered superior performance, efficiency and reliability in tasks involving the classification of medical images. Further study of this work can be carried out, optimizing model accuracy, robustness, and computational efficiency by including varied and more diversified datasets. Parallely, exploring novel architectures that can handle more complicated and obscure medical images opens up new areas of interest for upcoming works.