Beyond Accuracy: AI Brain Tumor Detection with GradCAM++ Interpretability & Clinical Deployment
CWSF · 2026 Digital Technology
Overview
Brain tumors, including glioma, meningioma, and pituitary tumors, are diagnosed too late, too inconsistently, and too far from where most people live; AI was supposed to fix all three. It hasn't, because clinicians cannot trust answers they cannot verify. I trained 40 deep learning models across four data augmentation strategies; no augmentation, basic, extreme, and domain-specific, to test whether each one attended to the correct anatomical region, not just whether it produced the correct label. The best reached 99.69% accuracy. But statistical validation (ANOVA F=38.9, p<0.001; Cohen's d=4.75) and GradCAM++ interpretability exposed a flaw in how medical AI is judged: two models with equivalent accuracy can attend to entirely different brain regions, a silent failure invisible to any accuracy score, and a 33× gap in validation loss. I then built IllumaDx, a deployable bilingual AI diagnostic system with GradCAM++ visualization that surfaces model reasoning to clinicians before they trust it.
Awards (2)
- Special Award
- Selected for CWSF 2026
Competition history
- CWSF 2026
Related projects
ISEF · 2025
An Analytical Study of Optimized EfficientNet CNNs, Vision Transformers, and Hybrid Transformers for State-of-the-Art MRI-Based Brain Tumor Diagnosis With Integrated Explainable AI
ISEF · 2025
An Adaptive Grad-CAM Thresholding Technique to Improve AI Interpretability in MRI-Based Prostate Cancer Risk Stratification
ISEF · 2025
A Novel Convolutional Neural Network to Detect and Classify Brain Tumors
ISEF · 2023
NueROX: A Visual Diagnosing System Using a Fully Optimized Convolutional Neural Network Architecture for Rapid and Efficient Classification of Tumorous and Non-Tumorous Brain MRI's and a Model for the Segmentation, Dimension, and Severity Estimation of Brain Tumors
Closest projects by meaning, across every fair and year in the corpus.