Training a TensorFlow model to detect lung nodules on CT scans
summary
A Python TensorFlow pipeline trains a two-stage model that first generates voxel-wise candidate maps on 200×512×512 CT volumes (dataset ~5 TB) and then classifies patches; the localization stage targets ~97% sensitivity using dilated convolutions and a ResNet-like design, training is stabilized with weight normalization and heavy augmentation, and post-processing yields candidates re-scored to reach about 90% sensitivity at 1 false positive per scan versus ~80% for radiologists — a working system that tops a lung-nodule detection challenge.
AI-written summary · about
chapters (8)
- 0:00Lung nodule detection project
- 4:20Pipeline and pre-processing
- 6:46Localization network
- 13:59Weight normalization and augmentation
- 17:39Post-processing and false positives
- 19:29Results on LUNA challenge
- 20:17Kaggle Data Science Bowl
- 24:14Future of deep learning
related talks
24mAutomating Porosity Detection with Deep Learning
55mDeep Learning in Medical Image Diagnostics
43mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
25mDeep Learning in Biomedicine: Challenges and Opportunities
57mTensorflow, deep learning and modern convolutional neural nets, without a PhD
42mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
49mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
15mBrain Tumor Detection and Classification Using Deep Learning Models