Code and Files

The description of each files are listed below. Due to privacy reason, data files are not listed in the repository:

Preprocessing

1. read dicom.py Read all .dicom files and extract information from DICOM files (date- time, type, subject ID, dimension etc.). Output as a table with list of folders and corresponding information.

2. read birads.py Clean the BIRADS data from folder table generated by read dicom.py. Filter the data with valid image and text information.

3. read biopsy.py Combine the biopsy result table with DICOM file according to the subject ID and date of completion. Generate and validate the final labeled table for each scan.

4. filter list.py Filter the specific sequence of MRI scans (T1, T2, etc) according to the key words written in the DICOM file.

FeatureExtraction

extract features.py Feature Extraction Agent. Extract features of DICOM files from pre- trained Inception v3 CNN. It automatically download and run the models on MRI images stored as DICOM files in a folder (1 scan per folder). The output of each scan is the feature map matrix n*2048 where n is the number of DICOM files and 2048 is the number of feature extracted.

Training-and-Testing

1. funclib.py Customized function library for training and testing.

2. rnn reinforce v3 training.py Train the Selection Agent and Prediction Agent using the reinforcement learning describe in Chap 4. The training is deployed on the CPU machine with tensorflow. Change of loss function and training accuracy are stored when training.

3. Inference.py Inference the result of trained Selection Agent and Prediction Agent. ROC curve and confusion matrix are generated during the inference to evaluate the performance of trained model.