This study introduces an innovative deep learning pipeline designed for precise medical image classification, specifically targeting the discrimination between "PUJ obstruction" and "Normal" conditions. By harnessing transfer learning from pre-trained CNN models such as VGG16, InceptionV3, and DenseNet121, the framework optimizes feature extraction for improved classification accuracy. Custom convolutional layers further enhance model performance, with rigorous evaluation conducted on labelled datasets using metrics like accuracy and F1-score. Advanced visualization techniques, including t-SNE for feature embeddings and activation maps for interpreting learned representations, offer novel insights into the model's decision-making process. This holistic approach represents a significant advancement in medical image analysis, promising to elevate diagnostic precision and ultimately improve patient care. Keywords— PUJ Obstruction, Normal, CNN, VGG16, InceptionV3, DenseNet121, F1-Score and t-sne This research introduces a approach that merges real-time ultrasound imaging with advanced machine learning to automate the detection of Pelvic Uretero Junction (PUJ) obstruction. Traditional diagnostic methods for PUJ obstruction often involve invasive procedures, leading to patient discomfort and risks. In response, this study aims to revolutionize diagnosis by improving accuracy and timeliness through the proposed system. By integrating ultrasound imaging with machine learning, the research seeks to provide a patient-friendly and efficient alternative to traditional diagnostic approaches, ultimately enhancing patient outcomes and facilitating earlier interventions.
