Jayy268/Brain_Tumor_Detection

Deep Learning for Medical Imaging: Comparing Custom CNNs vs. MobileNetV2. Validated on real-world BraTS-Africa data to prove robustness against domain shift # Deep Learning for Brain Tumor Detection: Addressing Domain Shift in Low-Resource Settings ## 🏥 Project Overview This project investigates the critical challenge of **domain generalization** in medical AI. It compares two deep learning approaches—a custom Convolutional Neural Network (CNN) and a Transfer Learning strategy (MobileNetV2)—to detect brain tumors in MRI scans. Beyond standard validation, this study rigorously tests both models on an **external, real-world clinical dataset (BraTS-Africa)** to evaluate performance under domain shift. The goal was to determine which architecture offers the safety and robustness required for clinical deployment in diverse environments. ## ⚡ Key Highlights * **Dual-Path Experimentation:** Benchmarked a Custom CNN (trained from scratch) against a frozen MobileNetV2 (Transfer Learning). * **Real-World Validation:** Utilized the **BraTS-Africa** dataset (30 patients, ~150 slices) as a completely unseen external test set to measure robustness. * **Critical Discovery:** Uncovered a severe case of **"Shortcut Learning"** in the custom model (0% external sensitivity) vs. robust generalization in the transfer learning model (79.3% external sensitivity). * **Clinical Metrics:** Prioritized **Sensitivity (Recall)** and **Specificity** over raw accuracy to reflect clinical safety requirements. ## 📊 Methodology ### The Two Approaches 1. **Path 1: Custom CNN** * A lightweight, 4-layer Convolutional Neural Network built from scratch. * Designed to test the efficacy of learning features purely from a small medical dataset (~200 images). 2. **Path 2: Transfer Learning (MobileNetV2)** * A state-of-the-art architecture pre-trained on ImageNet. * **Strategy:** "Frozen" feature extraction. We froze the base layers to leverage robust, generalized visual filters and only trained the classification head. ### The Datasets * **Internal Training/Validation:** Public Kaggle Brain Tumor MRI Dataset (Standardized, clean …