DICOM Image Parsing and Training Pipeline

Programming Language

• Python 2.7

Library Dependencies

• numpy r1.13.0
• pandas r0.20.2
• dicom r0.9.9
• opencv r3.2.0
• matplotlib
• PIL

Parsers (file: parsers.py, testing file: test_parsers.py)

• DicomParser: a class for parsing DICOM image
• ContourParser: a class for parsing contour files and rendering a contour polygon
• MaskParser: a class for producing a boolean mask given DICOM image and contour polygon
• test_parser.py: to run unit tests on parsers.py, simply run $ python test_parsers.py on terminal

Training Pipeline (file: pipeline.py, testing file: test_pipeline.py)

• TrainingPipeline: a class for pairing DICOM images and contour files, producing boolean masks, parsing all data, and batch serving (input, target) pairs
• test_pipeline.py: to run unit tests on pipeline.py, simply run $ python test_pipeline.py on terminal

Demos (file: demo1.ipynb and demo2.ipynb)

• demo1.ipynb: an out-dated demo of the main APIs of parsers.py and pipeline.py is illustrated in the demo.ipynb notebook. In addition to the demonstration of the major APIs, a few discussions on how the pieces are developed, how the correctness of the program is verified, and how the parser and pipeline can be further improved in the future are also included.
• demo2.ipynb: an up-to-date demo of the main APIs that both parsers.py and pipeline.py offer.

Heuristic LV Segmentation Analysis (file: explore_heuristic_LV_segmentation.ipynb)

• In this notebook, a few heuristic approaches for LV segmentation are explored, compared, and analyzed, including a) a universal fixed single-value threasholding scheme; b) a case adaptive threasholding scheme; c) watershed algorithm for image segmentation.