ml-w/pytorch_medical_imaging

[WIP]

Introduction

This repository aims to be a pipeline that uses Pytorch to train and inference deep learning model for medical imaging data.

Installation

TLDR;

Getting the source code:

git clone --recursive https://github.com/alabamagan/pytorch_medical_imaging
git clone https://github.com/alabamagan/mri_normalization_tools

Install custom repos that are pre-requisits:

pip install ./mri_normalization_tools
pip install pytorch_medical_imaging/ThirdParty/torchio # forked version refined for this package
pip install pytorch_medical_imaging/ThirdParty/surface-distance # for in-built system to evalute performance

Install the main package locally:

pip install pytorch_medical_imaging

Third Party Packages (customized forks)

Torchio

This repo uses mainly torchio as the IO, however, as torchio lacks certain function we require, we forked the repository and made some changes that are accustomed to our needs here.

Alternative, you can install the forked package using this command:

pip install git+https://github.com/alabamagan/torchio

MRI image normalization tools

This package uses the logger from MNTS, which is the normalization tool I wrote for convinience and reproducibility.

To install:

pip install git+https://github.com/alabamagan/mri_normalization_tools

Specification

pmi is implemented with 4 main units which interacts for training and inference:

1. main
2. pmi_data_loader
3. pmi_solver
4. pmi_inferencer

Model training

sequenceDiagram 
	Actor User
	User ->> main: Configurations
	main ->> pmi_data_loader: Data load config
    loop
        pmi_data_loader ->> pmi_data_loader: Load train data
        pmi_data_loader -->> pmi_data_loader: Load validation data
    end
    main ->> pmi_solver: TB plotter for visualization
    rect rgb(15, 55, 35) 
        loop
            pmi_data_loader ->>+ pmi_solver: tio.Queue
            pmi_solver ->> pmi_solver: Training epochs
            pmi_solver -->> pmi_solver: Validation
	        pmi_solver ->> main: Return data if requested
        end
	end
    pmi_solver ->>- main: Training finish, return learnt states
    main ->> User: Saved states

Loading

Model inference

sequenceDiagram 
	Actor User
	User ->> main: Configurations
	main ->> pmi_data_loader: Data load config
    loop
        pmi_data_loader ->> pmi_data_loader: Load inference data
        pmi_data_loader -->> pmi_data_loader: Load target data

    end
  	main ->> pmi_inferencer: Trained state
    rect rgb(15, 55, 35) 
        loop
       		pmi_data_loader ->>+ pmi_inferencer: tio.Queue
            pmi_inferencer -->> pmi_inferencer: Inference
            pmi_inferencer ->> pmi_inferencer: Compute performance
            pmi_inferencer ->>- User: Write results to disc
        end
	end

Loading

Call hierachy

%%{ init : { "theme" : "dark", "flowchart" : { "curve" : "linear"}}}%%
flowchart TD
	subgraph configs
		sc(Solver cfg)
		lc(Data loader cfg)
		sc & lc --> pmic(Controller cfg)
	end
	N[Network] --> sc
        N --> B
	pmic --> |populate|B[PMI Controller]
	B --> |create|dl[Data loader] --> Solver["Solver/Inferencer"]
	B --> |create|Solver
	Solver --> |call|fit[/"fit()/write_out()"/]

Loading