Highly performant breast lesion malignancy detection and ACL tear detection models built using transfer-learning of large CNNs. Interpretable model decisions powered by Grad-CAM.

(Arxiv Link Pending)

experiment.py is the most comprehensive wrapper script for our analysis. With --method runall, this manages gridsearching for optimal architectures by passing cross-products of hyperparameters and architecture choices to main(). With --method summarize --filter key value the script produces a results/results.csv file containing all metrics for each hyperparameter combination where key = value (e.g. if we set --filter epochs 10, we will retrieve all results for experiments run with --epochs 10). Note that each row in the results.csv corresponds to the 'best'/checkpointed epoch from model training, as selected by highest validation AUC. With --method visualize the script produces a set of visualizations in results/acl/, results/breast/, and results/overall/ which provide box-plot + experimental scatters comparing hyperparameter choices' affect on metrics.

Example (also triggerable as a sequence with ./run_experiment.sh):

# Run all experiments defined in experiment.py loops
python experiment.py --method runall --verbose

# Summarize all epoch=10 experiments into a single CSV
python experiment.py --method summarize --verbose --filter epochs 10

# Create visualizations from the summarized results
python experiment.py --method visualize --verbose

main.py handles dataloader setup and device setup, and serves as a point of contact for users to trigger new experiments from the CLI (and for experiment.py to start grid search experiments).

Examples (Best Breast Model and Best ACL Model):

$ python main.py --data_dir breast --database ImageNet --backbone_model_name ResNet50 --clf ConvSkip --structure unfreezetop5 --verbose --dropout_prob 0.5 --fc_hidden_size_ratio 1.0 --num_filters 16 --kernel_size 2 --epoch 30 --batch_size 64 --lr_decay_method cosine --amp --lr 5e-4

$ python main.py --data_dir acl --database ImageNet --backbone_model_name ResNet50 --clf ConvSkip --structure unfreezetop5 --verbose --dropout_prob 0.5 --fc_hidden_size_ratio 0.5 --num_filters 16 --kernel_size 4 --epoch 30 --batch_size 64 --lr_decay_method cosine --amp --lr 1e-3

interpret.py handles all Grad-CAM logic for generating and visualizing Grad-CAM heatmaps to interpret model results.

Example:

$ python interpret.py --data_dir breast --database ImageNet --backbone_model_name ResNet50 --clf ConvSkip --structure unfreezetop5 --verbose --dropout_prob 0.5 --fc_hidden_size_ratio 1.0 --num_filters 16 --kernel_size 2 --epoch 30 --batch_size 64 --lr_decay_method cosine --amp --lr 5e-4 --image_index 0

predictions.py is a simple script for producing predictions/preds_{MODEL_PARAM_STR}.csv files with all the test predictions for a particular model.

Example:

$ python predictions.py --data_dir breast --database ImageNet  --backbone_model_name ResNet50 --clf ConvSkip --structure unfreezetop5 --verbose --dropout_prob 0.5 --fc_hidden_size_ratio 1.0 --num_filters 16 --kernel_size 2 --epoch 30 --batch_size 64 --lr_decay_method cosine --amp --lr 5e-4

src/ contains the source code for argument parsing, dataloader setup, model architecture building in PyTorch, and other utils.

data/ contains the data for all of the downstream classification tasks. We focus primarily on data/breast/ and /data/acl/. Each of this subdirectories contains folders datafram/e, images/, and models/. The dataframe folder contains the five-fold splits used by RadImageNet, as well as combined, re-split 75/15/10 train/val/test stratified (on target) splits that we generate and use. Each row contains a label and an image path, which points to an image in images/. models/ contains training histories (performance metrics throughout training) as well as checkpointed models, though much of this is not uploaded to github due to filesize constraints.

logs/ is used for TensorBoard logging, and should also be mostly empty on github.

predictions/ contains predictions for our best breast and ACL models, as well as their less performant RadImageNet initialized counterparts.

results/ contains gridsearch and unfreezing experiment results and visualizations.

tflow_replicated_expts/ contains debugged code from the original RadImageNet repo, used to compare results for our Linear baselines models.

Updates History:

PyTorch v3 had fixes to Caffe preprocessing, train dataloader shuffling (especailly important for ACL), and a handful of other fixes.

PyTorch v4 architecture removes the softmax from the classifier appended to the backbone, relying instead on SoftmaxLoss so that we don't do a double softmax. This massively improves breast performance.

After Refactor May 29:

Example usage: python main.py --data_dir acl --database RadImageNet --backbone_model_name ResNet50 --clf NonLinear --structure freezeall --verbose --dropout_prob 0.5 --fc_hidden_size_ratio 0.5 --num_filters 8 --kernel_size 2 --epoch 5 --batch_size 64

See main.py for the full list of arguments. model.py handles training the models, as well as defines the Backbone and Classifier layers. util.py validates arguments and provides functions for loading data. main.py parses arguments, sets device, and iterates through training and validation folds.

====== 05/31: ====== Aditri added Convolutions with Skip Connections as an option. Daniel added data prep options to run against full train/val/test splits and re-split and aggregated the data to ensure no leakage.

Daniel added LR scheduling, more dynamic model checkpointing for all hyperparameters.

Daniel added linting.

TBD: Daniel adding SWA, experiment.py for running a vast grid of experiments + summarizing experiments into overall results/results.csv, visualizations for report.