Discriminating solid masses and cysts in dual-energy digital mammography using a convolutional neural network: Simulation study
Records and experiments of training deep learning model to differentiate solid masses and fluid cysts in dual-enery mammograms using Monte-Carlo simulation programs developed by the FDA with anthropormorphic digital breast phantoms supplied by Christian Graff and mentored by Stephen J. Glick and Andrey Makeev from OSEL/CDRH/DIDSR.
The model weights which achieved 97.02% validation accuracy (trained for 50 epochs) is located in the best_try folder in the file called best_try27.h5. A ResNet architecture was used using (from PyImageSearch) Adrian Rosebrock's implementation with 32 layers and three stages of filters with sizes 64, 128, 256, and 512 (resnet.py).
The Python program modelmk1.py is the file we ran to train the model and it produced files saved with the weights from the best validation loss, located in the best_try folder, and the ending weights of the model after training, located in the ending_weights folder. Graphs of the training and validation loss and accuracy are saved in the plot folder and dictionaries containing the values are saved in the history folder in the form of pickle files. The monitor folder contains .json files and .png graphs which are updated every epoch during training in order to monitor the validation loss curve to ensure the model is not overfitting. In the case where it is, then we are able to stop the model and tune the right hyperparameters to reduce the time spent on training (implemented by Adrian Rosebrock from PyImageSearch). The file trainingmonitor.py is the callback implementation used and is located in the pyimagesearch folder. The onecycle folder includes the loss values and the learning rate values for one epoch of training in order to find the optimal max learning rate to use during cyclical learning rate one cycle training as well as corresponding plots of learning rate vs. training loss. The suggested value of the maximum learning rate should be the largest learning rate with lowest loss before the loss explodes or becomes unstable and the values can be generated by running find_lr.py which uses the LRFinder class in clr.py. The model then uses cyclical learning rates modified from fast.ai Jupyter Notebooks with two cosine functions.
Sample images for the validation and training set are located in the cyst and mass folders.