Please enable JavaScript.
Coggle requires JavaScript to display documents.
Methodology - Coggle Diagram
Methodology
Test set reviewing process
Initial 4-Dr.'s annotations
Methods to align annotations. Retrieval of LN consensus and metastasis consensus
Additional IHC stained slide re-slicing, scanning, labeling and finalize labels on equivocal LN
Statistical method to evaluate the results
Equivocal case review w/ AI's help
Dataset & Annotations for training
Fully-annotated slides for evaluating localization
How did we split datasets? Why sampling LKCGMH 2019 for the main test set?
LN-level annotations (semi-auto method)
Slide-level annotations
Data description
Slide retrieval and sampling from LKCGMH 2018, 2019 (full data), 2020, KHCGMH 2019, 2020, 2021
Staining, Scanning
IRB
Training workflow
Lymph node detector
DeepLabV3 to generate masks
Post-processing to convert masks to contours
Why not detection model: Cropping issue
Whole-slide training method for classification
Model architecture
Quick description of streaming CNN
Patch-based image augmentation technique to boost training performance
Use GPU to speed up image augmentation
Challenge: GPU memory constraint
Patch-based image augmentation and partial image rotation method
Bottleneck on image processing
Clinical trial to evaluate the improvement receiving assistance from AI
Experiment design: tasks, assignment, phasing, ...
Check time statistical analysis
Annotation integrity analysis
Cross-site evaluation
Direct evaluate
Finetune and re-evaluate