Skip to content

Readme.md

Latest commit

 

History

History
37 lines (35 loc) · 5.03 KB

Readme.md

File metadata and controls

37 lines (35 loc) · 5.03 KB

License

Ophthalmo-CDSS by Robert Andreas Leist is licensed under CC BY-NC-SA 4.0

Backend Documentation

Overview

This document provides an overview of the backend of the OphthalmoDashboard. The backend is responsible for the communication between the frontend and the data. It is implemented in the src/ folder of the repository. The backend consists of several parts:

  • data/: Functionality concerning data base access, preprocessing and loading.
  • prognosis/: Time series forecast model architectures, weights and training scripts.
  • segmentation/: YNet architecture and weights. This folder is a mirror of the YNet repository.
  • alignment.py: Contains the code for aligning the OCT images.
  • recommendation.py: Contains the code for the recommendation system.
  • reconstruction.py: Contains the code for the reconstruction of OCTs.
  • util.py: Contains utility functions used by the backend.
  • visualisation.py: Contains functions for visualizing the data.
  • vol.py: Contains a VOL class for loading OCTs in the VOL format. This class was adapted from the Heidelberg Engineering Heyex VOL reader. It features an automatic segmentation of the OCT and computes three-dimensional reconstructions from segmentations.

Prognosis

The prognosis module contains the time series forecast model architectures, weights and training scripts. The module contains the following files:

  • architectures.py: Contains the model architectures for the time series forecast. Implemented architectures are:
    • BasicRNNRegressor: A simple RNN model.
    • BasicLSTMRegressor: A simple LSTM model. Consisting of configurable LSTM layers and a fully connected layer.
    • AdvancedLSTMRegressor: Slightly advanced LSTM model. Consisting of configurable LSTM layers, a layer normalization layer and a fully connected layer.
  • models.py: Contains convenience classes to combine models of varying time targets. Classes also automatically load weights from the prognosis/weights/ folder.
  • train.py: Contains the training script for the time series forecast models.
  • weights/: Contains the weights for the time series forecast models.
    • Contains a subfolder for each predicted metric n_fluids, v_fluids and visual_acuity.
      • Each folder named after the predicted metric contains subfolders, which contain weights for different time targets of the forecast. The naming convention for the subfolders is "interpolation_method_metric_time_target".
        • The weights are stored as .pt files. The naming convention for the weights is "model_name_parameters.pt". For example: AdvancedLSTMRegressor_hidden_16_layers_2_lr_0.01_redf_0.9_batch_128_loss_L1Loss_drop_0.1.pt is an AdvancedLSTMRegressor model with 16 hidden layers, 2 LSTM layers, a learning rate of 0.01, a reduction factor of 0.9, a batch size of 128, a loss function of L1Loss and a dropout rate of 0.1.
      • Each folder also contains the results of a training run for the different time targets. The naming convention here is "train_results_time_target.csv". The tables contain the training and validation loss and hyperparameters for different models and are created by the train.py script.
    • Computer StandardScalers are also saved in this folder. They are used to scale the input data for the models and inverse scale the output. Their naming convention is "interpolation_method_metric_time_target_x or y depending on whether features or target should be scaled_scaler.pkl".

Segmentation

The segmentation module contains the YNet architecture and weights. This folder is a mirror of the YNet repository. Only a folder with the weights is included in this repository. The weights are stored in the segmentation/pretrained_models/ folder. The weights are stored as .pt files. Convenience classes for loading the weights are provided in the segmentation/__init__.py file.