Model-in-the-loop: A Pipeline for Same-Session Validation of Retinal Digital Twins for Calcium Imaging

This repository accompanies my MSc Thesis.
It builds on djimaging, open-retina (see D’Agostino, Zenkel, et al., 2025), and QDSpy. Parts of the code are copied from or modified versions of code found in these repositories.

The goal of this repository is to enable semi-automatic closed-loop experiments in which digital twins can be trained and tested within a single experiment session.

---

Installation

1. Clone the following repositories:

   • djimaging
   • open-retina
   • ScanM support

2. Update the paths in the requirements.txt file so that they point to the cloned repositories.

3. Create an environment and install the required packages.

4. Adjust config.yaml. The following fields must be updated:
   repo_directory, home_directory, dj_config_directory, recording_files_dir, stimulus_output_dir, set_cache_dir_openretina, load_model_path.
   See the comments in the file for guidance.

5. Update the username in config/DJ.yaml. You can also adjust parameters of the djimaging tables here.

---

Description

core/
Contains DJComputeWrappers, schemas, and GUI-related code.

data/
Directory where new data should be placed (data/new_data_dump).

models/
Pretrained models used for refinement.

utils/
Helper scripts, including:

• transform_to_avi_stimulus.py: functions for converting optimized stimuli to .avi, saving stimuli and metadata in an iteration folder.
• model_trainig.py: utilities for refinement (training-related helpers).
• QDSpy_helpers/: functions imported by the Stimulus PC for reading metadata.