Navigating Prevalence Shifts in Image Analysis Algorithm Deployment

This repo serves as reproduction code for the following papers

• "Deployment of Image Analysis Algorithms under Prevalence Shifts" see Springer and ArXiV. It was published as conference paper at MICCAI 2023.
• "Navigating Prevalence Shifts in Image Analysis Algorithm Deployment" (unpublished) an extended version of the former, submitted to the Medical Image Analysis journal.

In the following, the reproduction descriptions only focus on the latter, extended version. To reproduce our original results please use the respective git tag and check out "miccai23".

Overview

• code structure
• installation
• image data preparation
• model training and prediction generation
• experiments and figures

code structure

The code of this repository is structured as follows in src:

• mml_plugin implements a mml plugin to re-distribute samples within a task according to our needs
• prev contains definitions and routines that are shared through our experiments (note that __init__.py modifies psrcal behaviour)
• training_scripts contains all commands with respect to task data (prepare, preprocess) and neural networks (train and predict)
• the notebooks 1_... to 8_... contain the steps to reproduce our experiments

pre-requisites

To run the notebooks and reproduce the plots exactly you might need to install the font we used.

• download NewCM10-Regular.otf
• place it in the /data folder at project root (necessary for notebook 5)
• install the font (the details of this step depend on your OS)

mml-free reproducibility

mml was primarily used to handle the imaging datasets, run the model training and produce the prediction logits for this project. All analysis on top can be done without mml, which we explain here:

• create a virtualenv with conda and install python 3.10, install the requirements

conda update -n base -c defaults conda
conda create --yes --name prev python=3.10
conda activate prev
git clone https://github.com/IMSY-DKFZ/prevalence-shifts.git
cd prevalences
pip install -r requirements.txt

• Run the following notebooks to follow our experiments:
  • 3_prevalence_estimation.ipynb - Research Question 1, creates figures 5 and C.11
  • 4_calibration.ipynb - Research Question 2a, creates figure 6
  • 5_threshold_visualization.ipynb - Research Question 2b, creates figures 3, 7, and 8
  • 6_decision_rule.ipynb - Research Question 2b, creates figure 9
  • 7_validation_metrics.ipynb - Research Question 2c, creates figure 10
  • 8_uncertainty.ipynb - Creates uncertainty tables

installation

These instructions handle the full reproduction including the installation of mml (see above for a light setup).

• create a virtualenv with conda and install python 3.10

conda update -n base -c defaults conda
conda create --yes --name prev python=3.10
conda activate prev

• install mml-core and mml-tasks plugin

pip install mml-core
pip install mml-tasks

• install local prevalence plugin and other requirements

git clone https://github.com/IMSY-DKFZ/prevalence-shifts.git
cd prevalences
pip install -r requirements.txt
cd src/mml_plugin/prevalences
pip install .

• setup environment variables for mml

cd ../../..
mml-env-setup
nano mml.env  # modify at least MML_DATA_PATH, MML_RESULTS_PATH and MML_LOCAL_WORKERS accordingly 
pwd | conda env config vars set MML_ENV_PATH=$(</dev/stdin)/mml.env
conda activate prev

image data preparation

Note: We do not own the datasets used for this study and many datasets have to be downloaded manually. Instructions will be given along the "01_create_cmds" process.

• the data and predictions generation process in handled with the mml framework
• the commands to leverage mml are generated in 1_generate_predictions.ipynb and stored in training_scripts
  • 01_create_cmds.txt for data download / task generation
  • 02_pp_cmds.txt for data preprocessing
  • 03_tag_cmds.txt for the splitting of tasks according to our experimental setup (train-validation-development test and deployment test)
  • 05_dataseed_cmds.txt for creating 5 additional splittings with different splitting seeds
• if the commands shall be run on some external infrastructure (like a GPU cluster) the 1_generate_predictions.ipynb contains configuration possibilities to adapt the txt files
• the commands can be run locally by bash 0X_XXX_cmds.txt (stick to the order indicated by numbering)

training and predictions

Note: Model training requires appropriate hardware (e.g. a GPU), ideally on a distributed cluster.

• once more mml is leveraged for this step and the commands are generated in 1_generate_predictions.ipynb and stored in training_scripts
  • for uncertainty assessment (8_uncertainty.ipynb):
    • 04_reproduce_cmds.txt for re-training the original experiments with 6 additional seeds on the original splits
    • 06_dataseed_predict_cmds.txt for re-training each task once per additional splittings
  • for re-calibration assessment (4_calibration.ipynb):
    • 07_retraining_cmds.txt for re-training on the original splits with adapted loss weights according to exact deployment prevalences (for each IR 1.0, 1.5, ..., 10.0)
    • 08_retraining_estimated_cmds.txt for re-training on the original splits with adapted loss weights according to estimated deployment prevalences - using ACC (for each IR 1.0, 1.5, ..., 10.0)
• the files can be run or adapted as mentioned before
• keep in mind that they incorporate (6+5+19+19) * 30 = 1470 training+prediction pipelines and take some time to complete
• in our analysis we also used the original training and previous 3 additional seeds next to 06_dataseed_predict_cmds.txt (in sum 10 seeded repetitions on the original split)

experiments and figures

DISCLAIMER: For licensing reasons we may not provide all predictions, but attach some sample predictions in /data. We also provide intermediate results:

• 24_prev_estimation_df.pkl - holds all quantification results (RQ1)
• 24_recalibration_results.csv - holds all re-calibration results (RQ2a)
• 24_decision_rule_results_....pkl - holds results on applying various decision rules (RQ2b)
• 24_metric_performance_....pkl - holds all metric evaluations (RQ2c)

• navigate to the top level folder named data and store the project folders generated by the previous commands in there, more precisely
  • locate your MML_RESULTS_PATH as provided in the installation
  • within search the project folders
    • original publication (n=4):
      • mic23_predictions_original_0
      • mic23_predictions_reproduce_0, ..., .._2
    • generated by 04_reproduce_cmds.txt (n=6):
      • mic23_predictions_original_10, ..., .._15
    • generated by 06_dataseed_predict_cmds.txt (n=5):
      • mic23_predictions_datasplit_seed_3, ..._31, ..._314, ..._3141, ..._31415
    • generated by 07_retraining_cmds.txt (n=19):
      • mic23_predictions_extension_balanced_0_1.0, ..._0_1.5, ..._0_2.0, ..., ..._0_10.0
    • generated by 08_retraining_estimated_cmds.txt (n=19):
      • mic23_predictions_extension_balanced_estimated_0_1.0, ..._0_1.5, ..._0_2.0, ..., ..._0_10.0
  • copy those to the data folder next to src
• now you can load the predictions inside the jupyter notebooks, they should run straight forward and produce the figures in results, next to src and data

Notebooks:

• 2_plot_examples.ipynb - Task overview, Figure 4
• 3_prevalence_estimation.ipynb - Research Question 1, creates Figures 5 and C.11
• 4_calibration.ipynb - Research Question 2a, creates Figure 6
• 5_threshold_visualization.ipynb - Research Question 2b, creates Figures 3, 7, and 8
• 6_decision_rule.ipynb - Research Question 2b, creates Figure 9
• 7_validation_metrics.ipynb - Research Question 2c, creates Figure 10
• 8_uncertainty.ipynb - Creates uncertainty tables