mimir-biomedical / mobile-unetplusplus-cac-scoring

Implementation of UNET++ for CAC Scoring using Tensorflow

Mobile UNet++ Model for Coronary Artery Calcium Scoring

THIS REPO IS STILL UNDER DEVELOPMENT, IMPROVEMENT IS WELCOMED!

A UNet++ with MobileNetv2 block based ML model to do automatic CAC scoring This is a repo to develop my Undergraduate Thesis on Biomedical Engineering

Development Environment

• WSL2 Ubuntu Linux 20.04.4 LTS
• Miniconda (for environment management)
• Poetry (for dependencies management)
• Python 3.10.8

Library

• Tensorflow
• Streamlit
• PyDicom
• Numpy

Reference

• Paper:
  • Original Paper using COCA Dataset (they use UNET Architecture)
  • Calcium Scoring Method
• Code:
  • keras-unet-collection
  • azcopy install script

Dataset

The dataset used for this model is provided by Stanford AI COCA Dataset, this dataset can be accessed through this link

This model is using Gated MRI Data from the COCA Dataset consisting of the MRI image saved in a DICOM format and the calcium mask saved in a xml plist format.

Other dataset usage will be developed after the implementation of the COCA Dataset has been done.

mobile-unetplusplus-cac-scoring's Issues

Model: Mobile UNet with multi head

Model: Custom Loss (Log Cosh Dice Loss)

Compress and shard tfrecord data

Add weighted categorical crossentropy

Ensure dependencies can work with python 3.8

Preprocessing-Segmentation: Clean python dictionary

Postprocessing: Connected Component Labelling

Model: Data interface with model

Fix tfrecords data loader for training

Troubleshoot exploding/vanishing gradient

Add gcloud script to upload and download processed dataset

Add notebook to test model

Fix issue due to noninversed binary segmentation used for multiclass segmentation

Fix naming in model using non deepsupervision

Add model size to metadata

Remove binary segmentation as model output

Try every possible loss function

Builder for both UNet++ and MobileUNet++

Fix cleaning raw segmentation: make left circumflex artery to abbreviate to lcx

Preprocessing-Segmentation: Convert plist to python dictionary

Added requirements.txt from pyproject.toml

Refactor Model Builder

Preprocessing-DICOM: Get CT Scan Image and HU representation

Add null class on multiclass segmentation on create_dataset func

Fix default mobile config number of class

Revamp gcs-cli and make gcs-cli-non to use official cli

Postprocessing: Agatston Algorithm on post processing

Resample dataset to reduce imbalanced dataset

Fix model training cli

Add missing preprocess function on auto cac system

Create visualization figure for preprocessing steps

Model: MobileNetV2 Blocks

Rework preprocessing method

Mixed precision to reduce gpu memory usage on training

Fix cleaning raw segmentation data process to skip empty roi

Add flood fill for segmentation data to get full segmentation rather than perimeter segmentation

Feat: Added patient dataset blacklist preventing duplicate pixel location

Preprocessing-Segmentation: Split feature for binary and multiclass

Postprocessing: create postprocessing pipeline

Move flood fill algorithm on segmentation preprocessing

Preprocessing-DICOM: Add Pixel Spacing and Slice Thickness to extracted data

Refactor base model to be more universal

Non interactive mode on gcs-cli + default

Multihead UNet++

Model: Base UNet++ Model

Fix tfrecord generation

Preprocessing-TFRecords: Create TF Record per patient (contains binary and multiclass feature with its image)

Add CI for pytest

Add git hook for pytest

Breaking tasks down