Computer Vision in Melanoma Detection
summary
Uses public ISIC/HAM10k-style sets with strong class imbalance and common artifacts (rulers, hair, gel bubbles) that can leak labels; notes one dataset had rulers on every malignant case. Trains ResNet50 and Xception pretrained on ImageNet, stresses careful augmentation so ABCD features (asymmetry, border, color, diameter, evolving) survive, and boosts performance with localization — YOLO/object detection and spatial attention outperform naive crops and segmentation. Optimizes for Precision at recall 0.95 to minimize false negatives; ensembles on Kaggle beat single models.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:50Melanoma detection motivation
- 3:24Visual screening challenges
- 5:37Public datasets for research
- 9:32Dataset preparation challenges
- 11:20Deep learning modeling
- 14:12Data augmentation impact
- 16:23Localization and attention
- 19:16Model performance metrics
- 21:03AI impact on dermatology
- 22:04Q&A
related talks
15mBrain Tumor Detection and Classification Using Deep Learning Models
43mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
25mDeep Learning in Biomedicine: Challenges and Opportunities
49mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
55mDeep Learning in Medical Image Diagnostics
36mObject detection to Instance segmentation: Learn to apply several algorithms...
24mMachine Learning & Medical Imaging: The Future of Early Cancer Detection
36mSkin Cancer Detection using AI