Please enable JavaScript.
Coggle requires JavaScript to display documents.
Lung Cancer, Xu, Y., Hosny, A., Zeleznik, R., Parmar, C., Coroller, T.,…
Lung Cancer
Background
Most common cancer
Hard to diagnose even with advancement in medicine
Requires multiple follow-up and treatments
Traditional are more qualitative
AI provides quantitative measures
Effective detection
Accurate segmentation
Enhance patient care through better decision support
Lacks image dataset
Use of transfer learning
Single timepoint
However, tumors are dynamic
Use of CT scans from multiple follow-ups
Time-series data
Network Architecture
Use of ResNet CNN
Trained on imagenet
One CNN for each sets of training input
Total of four
Gated Recurrent Units
Recurrent layers
Fed with outputs from CNN
Allows learning in timeseries data
Transfer Learning
ResNet trained on ImageNet
Use of only pretreatment images
Used as baseline comparison
Statistical Analysis
AUC
Wilcoxon rank sums
Results and Discussion
Results are shown in text format
Baseline resulted to low performance
Addition of time series data improved AUC
Able to distinguish low and high mortality risk groups
The use of CNN achieved this result
Segmentation of regions
Prediction of pathologic response
Materials and Methods
Dataset
Two independent cohorts with different image sets
Dataset B
Treatments and surgery images
Used for testing
Dataset A
Treatment only and follow-ups
Used for training and testing
Image extraction and preprocessing
Use of CT scan
Center of image based on pre-defined seed
1month images
3 months images
Pre-treatment images
6 months images
Interpolated images
Use of three images slices to mimic 3D images
Image Augmentations
flipping
rotation
translation
deformation
Xu, Y., Hosny, A., Zeleznik, R., Parmar, C.,
Coroller, T., Franco, I., Mak, R. H., and
Aerts, H. J. Deep learning predicts lung cancer
treatment response from serial medical imaging.
Clinical Cancer Research 25, 11 (2019), 3266–3275.
Note: Combination of CNN and RNN could yield better results
Note: Research more on GRU