SAR 2020 - AI Masters Class

Class Schedule

Date / Time	Location
Sun 3/1 2:00-4:00 pm	Hibiscus Room

Slides

Requirements

All lessons are designed to run in Google Colab, which is a free web-based version of Jupyter hosted by Google. You will need a Google account (eg, gmail) to use Colab. If you don't already have a Google account, please create one in advance at the account sign-up page. You can delete the account when you complete the lessons if you wish.

Lesson Notebooks

Notebooks on Intro to Python, DICOM Basics, Classification, and Segmentation for AI Masters Class

• Lesson 01 - Intro to Python
• Lesson 02 - DICOM Basics
• Lesson 03 - Classification
• Lesson 04 - Segmentation

Lesson 01 - Intro to Python

Very basic introduction to Python and the Jupyter/Colab environment. If you've never visited the command line, start here.

Lesson 02 - DICOM Basics

Brief introduction to reading DICOM files including information from the DICOM header and the pixel array. Relies on pydicom and concepts from fastaiv2 medical imaging libraries.

Lesson 03 - Classification

Build a simple image classifier to identify a chest x-ray as either frontal or lateral projection. Based on fastai library and Deep Learning for Coders lesson.

Lesson 04 - Segmentation

Adapts Facebook Research's detectron2 library to read native DICOM and segment the liver on CT scans. We'll feed single-channel full bit-depth Hounsfield unit pixel data into detectron2 - DICOM and segmentations are from the recent Combined Healthy Abdominal Organ Segmentation (CHAOS) grand challenge.

Key modifications to detectron2 to enable reading DICOM natively:

• Custom data mapper to read the dcm file pixel_array and convert to Houndfield units.

  • Applies random brightness and random contrast run time augmentations.
  • Does NOT apply random flipping as horizontal flipping seems to lead to some liver-spleen confusion for the algorithm.
  • Sets mask type to "bitmask" from default "polygons"

• Custom get_liver_dicts referenced by DatasetCatalog.register function

  • Loads from _train, _val, or _test file lists
  • Saves path to .dcm file as dataset_dict["file_name"]
    • Default detectron data mapper simply opens the image file from file_name. Our custom data mapper will instead read the dcm pixel array, convert to Hounsfield units, and load the image array from that
  • Gets image height and width from dcm.pixel_array
  • Creates bounding box automatically from the given mask (via separate bbox function)
  • Converts binary PNG mask to RLE format using pycocotools library

• Modifies base configuration file for the trainer

  • cfg.SOLVER.IMS_PER_BATCH, BASE_LR, and MAX_ITER are good places to start tweaking to modify performance
  • cfg.INPUT.FORMAT = "F" tells the model to expect single channel float array instead of default "RGB"
  • cfg.INPUT.MASK_FORMAT = "bitmask" change from default "polygons"

• Extends the DefaultTrainer as LiverTrainer

  • To use custom data mapper in the build_train_loader and build_test_loader functions.
  • To use custom config file with modifications as above

• Replaces the DefaultPredictor class with LiverPredictor

  • Change torch.as_tensor to use single channel input

• Note that windowing is only done when displaying example images or overlaid inference results

  • Pixel data that are loaded into the network are NOT windowed since that would needlessly compress the dynamic range to the (presumed) detriment of the model's learning.
    • In this way windowing in python/ML applications means something slightly different than windowing at a diagnostic workstation. Radiolgists know but may not fully appreciate (and do not seem to communicate to computer scientists) that windowing at the workstation doesn't change the actual pixel values - an ROI will be the same regardless of the window/level setting applied since it is measuring the true HU, not the screen's pixel greyscale values. Windowing in python/ML, on the other hand, fundamentally changes the underlying pixel values.