Transfer learning for image recognition in healthcare...
summary
Experiments use ImageNet‑pretrained Inception v3 and ResNet plus a small custom CNN on four datasets (histopathology patches for cancer, retinal scans ~4k images, microscopy protein images <15k, and cats vs dogs) and report 4–10x speedup in epochs to reach comparable validation, average ~7x, and about ~4–5% absolute AUC improvement on medical test sets. Freezing early layers harms accuracy, indicating features not directly reusable, yet copying only per‑filter scaling (mean/variance) boosts convergence. The small custom CNN matches large nets trained from scratch.
AI-written summary · about
chapters (14)
- 0:00Transfer learning in healthcare
- 1:15What is transfer learning
- 3:03Challenges in medical imaging
- 3:46Experiment design for classification
- 5:34Model architectures and datasets
- 8:47Training setup and metrics
- 11:19Computational resources
- 12:03Results: pre-trained vs random
- 15:29Performance on test sets
- 16:54Feature reuse and freezing layers
- 19:31Visualizing feature clusters
- 20:14Weight scaling impact
- 21:39Key takeaways
- 23:28Q&A
related talks
46mTowards Understanding Transfer Learning with Applications to Medical Imaging
52mTransfer Learning and Finetuning Deep Convolution Neural Network
55mDeep Learning in Medical Image Diagnostics
56mTutorial on Deep Learning II
25mDeep Learning in Biomedicine: Challenges and Opportunities
25mEnsemble-X: Your personal strataGEM to build Ensembled Deep Learning Models for Medical Imaging
50mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
42mDeep Learning for Computer Vision