Artificial Intelligence for Medical Imaging Project

Group 4

• Roland Németh
• Nienke Muskee
• Yunda Qiu
• Amany El-Gharabawy

Description

Welcome to the GitHub repository hosting the code and experiment settings of Group 4 that was used to participate in the project phase of Artificial Intelligence for Medical Imaging and PI-CAI challenge. The followind documents describes what is needed to reproduce our experiments.

Prerequisites

The experiments heavily rely on the code provided by the hosts. Therefore, I would like to refer the reader to the corresponding repositories:

picai_baseline The picai_baseline repository contains scripts to train a UNet models that were used in our experiments and as such essential to reproduce our results.[1,2]

picai_unet_semi_supervised_gc_algorithm The picai_unet_semi_supervised_gc_algorithm repository contains scripts to prepare the docker container for submissions and the an already trained baseline model which we also utilized in our experiments.

picai_labels [1] Contains the label data and clinical data of the challenge.

PI-CAI challenge The challenge website for additional information.

Please follow the instructions on their corresponding web pages for proper installation manual.

Model Weights

The UNet model weights resulting from our experiments were too large to upload to GitHub. They are available here.

Structure of the repository

This repository contains multiple directories. The directories and subdirectories are described in the upcoming subsections.

code

This directory contains the 5 subdirectories and the postprocessing.py file.

postprocessing.py

This file contains the development code of the postprocessing experiment.

hp_exp_jobs

This directory contains .sh or .bash used to submit jobs on Snellius. Not all jobs were successful (due to errors) and therefore not all of them were reported about in the project report. However, the ones not listed below can be submitted and reproduced on Snellius by just submitting the corresponding .sh or .bash file.

Experiment 5 were submitted but was cancelled after being deemed redundant. Experiment 6 took to long to complete with the queue time and therefore was cancelled. Experiments 7, 8, 9 and10 resulted in an error which we did not manage to debug. (The bug only persists on Snellius.) Experiment 12 contained SparseAdam optimizer which was not suited for training this kind of neural network. Experiments 14 and 16 took too much memory and eventually stopped with a memory error. Experiments 17, 18 and LeakyRelu-BatchNorm were unsuccesful modifications of the model architecture which resulted in errors which we didn't have time to debug.

process_files

This directory contains the modified process.py files that were used for local testing and constructing docker containers for the postprocessing experiment and clinical data experiments. While most of the code was taken from picai_unet_semi_supervised_gc_algorithm, it contains our modifications as well.

training_extensions

This directory contains modifications to the training procedures such as changing the optimizer.

tabular_data

This directory contains the script used to train our clinical data models.

utils

This directory contains utilities such as log analyzer and testing the output of the clinical data models.

docker

This directory contains 3 subdirectories.

1. docker_hp is a subdirectory that contains (almost) all necessary components to build and export a Docker container of the chosen hyperparameter (hp) experiment. Please copy the weights of the experiments you want to dockerize into the weights folder. (All weights can be downloaded from here.)
2. docker_postprocessing is a subdirectory that contains (almost) all necessary components to build and export a Docker container of the postprocessing experiment. Please copy the baseline weights into the weights folder before dockerizing. (The weights can be downloaded from here.)
3. docker_tabular is a subdirectory that contains (almost) all necessary components to build and export a Docker container of the tabular experiment. Please copy the baseline weights into the weights folder and the selected tabular model files (without their directory) from AIMI_project/tabular_models into the docker/docker_tabular/tabular_models before dockerizing it. (The weights can be downloaded from here.)

logs

Contains the retrieved logs from training on Snellius. The most up to date can be found logs/logs/logs

tabular_models

Contains the directories of the models. Ridge and MLP were not submitted to the challenge. If the reader wants to reuse these, only copy the files (not the directories) as instructed in the previous section.

References

[1]A. Saha, J. J. Twilt, J. S. Bosma, B. van Ginneken, D. Yakar, M. Elschot, J. Veltman, J. J. Fütterer, M. de Rooij, H. Huisman, "Artificial Intelligence and Radiologists at Prostate Cancer Detection in MRI: The PI-CAI Challenge (Study Protocol)", DOI: 10.5281/zenodo.6667655

[2]J. S. Bosma, A. Saha, M. Hosseinzadeh, I. Slootweg, M. de Rooij, and H. Huisman, "Semisupervised Learning with Report-guided Pseudo Labels for Deep Learning–based Prostate Cancer Detection Using Biparametric MRI", Radiology: Artificial Intelligence, 230031, 2023. doi:10.1148/ryai.230031