Medical Image Processing with Deep Learning

Abstract

This project involved building a complete, end-to-end system for classifying medical images, with a specific focus on balancing high accuracy with the speed needed for real-world use. By using a ResNet-18 backbone and transfer learning, I was able to take the model from a theoretical experiment to a practical, working Streamlit app. A key part of the work was moving past 'black box' predictions; by integrating Grad-CAM, the model can actually show its reasoning by highlighting specific anatomical areas, such as the sellar region in pituitary scans. I made sure the methodology was rigorous, using stratified data splits and local optimization on an Apple M3 chip to ensure the results were both stable and reproducible. Ultimately, this work provides a reliable way to automate brain tumor detection while prioritizing Explainable AI (XAI), which is essential for building trust in healthcare technology.