Building AI That Learns and Adapts: A Case Study in MRI Diagnostics
summary
Project starts with 201 FCD MRIs plus ~100 normal and varied-pathology scans, preprocesses with brain tissue segmentation and two expert-inspired filters (junction and extension) to create a four-channel input (T1, FLAIR, junction, extension), trains a VAE-regularized U-Net variant, and evaluates explainable voxel-level segmentations. Baseline recall is ~70% with 54% precision; prospective test shows neuroradiologists detect 3/9 cases vs model 7/9 but the model yields many false positives (≈86). Rebalancing to 30 FCD /115 normal /120 pathology and applying elastic weight consolidation in continual learning keeps sensitivity (8–9/10) and raises specificity to ~70%.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:19Focal cortical dysplasias challenge
- 10:08MRI data collection
- 14:39Neural network for FCD detection
- 21:22Model degradation over time
- 27:45Continual learning importance
- 33:20Continual learning for FCD detection
- 40:44Results and clinical impact
- 43:15Continual learning in LLMs
- 47:50Future of continual learning
related talks
43mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
50mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
42mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
44mBuilding Adaptive AI for MRI Diagnostics: A Continual Learning Case Study
15mBrain Tumor Detection and Classification Using Deep Learning Models
34mWhy AI in the Medical Field Both Sucks and Rocks...
47mThe whole truth about using deep learning in medical imaging
44mAI in the Clinic: Prediction is Better Than Cure