GA Progression Modeling and Analysis using Deep Learning

Geographic atrophy (GA), an advanced form of age-related macular degeneration, is a leading cause of vision loss globally. Current approaches to monitor GA rely on manual examination of longitudinal imaging, which is time-consuming and subjective. This study presents a longitudinal tracking pipeline using AI-enabled segmentation and registration of fundus autofluorescence (FAF) images that allows for measurement of atrophic areas and visualization of disease progression.

This repo contains all code required to execute our GA progression monitoring pipeline from FAF imaging.

Acknowledgements

Some of the code in this project is adapted from other repositories, particularly the registration algorithm which uses weights from the LightGlue repo and vessel segmentation from vineet1992's vessel segmentation repo.

Dependencies

Make sure you have pyenv and poetry setup to run code. Clone the repo and execute the following:

pyenv virtualenv 3.10.4 ga_progression_modeling
poetry install

How to Run on your data

In order to run our pipeline on your data, we require the imaging dataset to be provided as a CSV file containing the columns outlined in the config section below.

Setup config

The config file in src/configs/config.yaml should store all the parameters for running the pipeline. The definitions of each parameter are provided below:

Note: Do not edit the paths in the models section if you want to run the exact same models for segmentation and registration.

Section	Parameter	Definition	Default
dataset	modality	Name of the image modality	Af
dataset	folder	Parent folder of csv data	data/
dataset	csv_path	CSV path relative to parent folder	dataset.csv
dataset	image_col	Column in CSV with image path	file_path_coris
dataset	ga_col	Column in CSV that contains or should contain the GA seg path	file_path_ga_seg
dataset	vessel_col	Column in CSV that contains or should contain the Vessel seg path	file_path_vessel_seg
dataset	patient_id_col	Column in CSV that contains patient ID	PID
dataset	laterality_col	Column in CSV that contains patient laterality (OD/OS eye)	Laterality
dataset	modality_col	Column in CSV that contains the modality name	Procedure
dataset	examdate_col	Column in CSV that contains the date of the image	ExamDate
dataset	size_x_col	Column in CSV that contains the X dimensionality of the modality image	XSlo
dataset	size_y_col	Column in CSV that contains the Y dimensionality of the modality image	YSlo
dataset	scale_x_col	Column in CSV that contains the scale factor conversion in the X direction for the modality image	Scale_X
dataset	scale_y_col	Column in CSV that contains the scale factor conversion in the Y direction for the modality image	Scale_Y
models	segmentation-ga-hist_eq	Applies histogram equalization to the image as a pre-processing step	True
models	segmentation-ga-weights	Path to the weights for segmentation model	src/weights/ga_seg/epoch=32-step=1023.ckpt
models	segmentation-vessel-hist_eq	Applies histogram equalization to the image as a pre-processing step	False
models	segmentation-vessel-weights	Path to the weights for segmentation model	weights/vessel_seg
models	registration-method	Register images using affine or tps (thin-plate spline)	affine
models	registration-input	Register over image space or vessel space	image
models	registration-reg2start	Register all images to baseline?	True
models	registration-tps	Optional lambda parameter for thin-plate spline (if using thin-plate spline for registration-method )	1.
models	registration-size	Size at which to register images	256
models	device	Device to run pipelines on	cuda:0
results	folder	Folder to create to save output. Will by default create a results/ folder containing this sub-folder	test

Run main.py

After setting all the parameters in the config file, simply run python src/main.py --config PATH_TO_CONFIG_FILE

Outputs

Running src/main.py will save output in the results folder specified in the config:

Main Output:

1. images_processed.csv : Processed data with segmentations and registrations saved in columns.
2. longitudinal.pptx: Powerpoint presentation with longitudinal image segmentations, registered videos, and growth curves.
3. sequential.pptx: Powerpoint presentation with each image segmentation indivitually shown.
4. pipeline_files/images_1.csv : Subset of the raw data csv, which contains all the rows for the specific modality.
5. pipeline_files/images_2.csv : Same as images_1.csv but with paths to the GA segmentation and Vessel segmentation included in columns.
6. pipeline_files/images_3.csv : Same as images_2.csv but with GA areas computed and added in as an extra column.
7. pipeline_files/images_4.csv : Same as images_3.csv but with paths to the registration parameters saved in the columns.
8. pipeline_files/segmentations/ga/: Contains all the segmentations of GA in the images.
9. pipeline_files/segmentations/vessels/: Contains all the segmentations of vessels in the images.
10. pipeline_files/registrations/registration_logs/: Contains all the log files for registrations.
11. pipeline_files/registrations/registration_params/: Contains all the parameter files for registrations. These are saved as .pth files.
12. pipeline_files/registrations/registration_keypoint_matches/: Contains visuals of the keypoint matches between images found during registrations.
13. pipeline_files/presentations/longitudinal/: Contains all metadata for generating longitudinal.pptx
14. pipeline_files/presentations/sequential/: Contains all metadata for generating sequential.pptx