Diabetic Retinopathy Classification Using DNN

Poster: Diabetic Retinopathy Classification Using DNN

For Running the script follow the instructions here or run the code directly using

Change the hyper parameters in constants.py

This project was exported from pycharm, environment based on python 3.8 was used. Make sure Source directory is diabetic_retinopathy

1. Input Pipeline
   • IDRID Dataset(Training set Images- 413, Test set Images-103)
   • Image Resize (to 256*256)
   • Image crop (Box crop)
   • Image Normalization
   • Class balancing
2. Model Architecture
3. Training Routine
4. Model CallBacks:
   • Check point Callback - For saving model at desired interval(epoch frequency)
   • Tensorboard Callback - For logging training stats,Profiling
   • CSV Logger Callback - To save training logs in a csv file
5. Training from a check point
   • Initial epoch here is the point at which the training was interrupted
6. Evaluation
   • Confusion Matrix
   • Classification Report
7. Data Augmentation
   • Vertical Flip
   • Horizontal Flip
   • Box Crop
   • Rotate
8. Deep Visualization
   • GradCAM
9. Hyper Parameter Tuning
   • Grid Search (Epochs, Number of dense neurons, stride, Learning rate)

Outputs from several stages of project

• After Image processing and data augmentation:

Processed and Augmented Images

• Model Architecture

  1. Model based on ResNET50v2 (Transfer learning).
  2. Model based on Blocks of Conv+BatchNorm+Maxpool (Only 453K parameters)

  

  Custom Architecture based on blocks of Conv+BatchNorm+Maxpool (ii)

• Training Results

  • Model based on Blocks of Conv+BatchNorm+Maxpool

  

  Accuracy plot for Custom model (on left) and ResNET50v2 based model (on right) [epochs vs accuracy]

• Results and Evaluation

  • Test accuracy - 77.8% (Custom Model), 81.55% (Finetuned on ResNET50v2)

  

  Confusion Matrix

  	precision	recall	f1-score	Support
  NRDR	0.73	0.82	0.77	39
  RDR	0.88	0.81	0.85	64
  accuracy	0.82	0.82	0.82	103
  macro avg	0.80	0.82	0.81	103
  weighted avg	0.82	0.82	0.82	103

  Classification Report

• Deep Visualization

  

  Original Image, GradCAM output, Overlay

• Hyperparameter optimization

  

  Original Image, GradCAM output, Overlay

Instructions to run the script:

Before running the script Install the requirments from requirements.txt using pip install -r requirements.txt

• Make the following changes in main.py based on the Mode(training mode, hyper parameter tuning mode, finetuning mode, evaluation mode) you want to run the script in.

  1. To Train the model, change the train FLAG in main.py to True
flags.DEFINE_boolean('train', True, 'Specify whether to train or evaluate a model.')

     • To log the data , specify path to tensorboard callback, model chekpoint call back, CSVlogger call back in constants.py dir_all_logs = 'log_dir'
dir_fit = os.path.join(dir_all_logs, 'fit')
dir_cpts = os.path.join(dir_all_logs, 'cpts')
dir_csv = os.path.join(dir_all_logs, 'csv_log')

  2. For performing hyperparamter Tuning, change the hparam_tune FLAG in main.py to True

     flags.DEFINE_boolean('hparam_tune', True, 'Specify if its hyper param tuning.')

  3. For Training the model based on ResNET50v2 change the Transfer_learning FLAG in main.py to True

     flags.DEFINE_boolean('Transfer_learning', True, 'to use transfer learning based model, train flag must be set to true to fine tune pretrained model')

  4. For Evaluating the pretrained model

     • Change the path of the pretrained model here in main.py to desired path.

     flags.DEFINE_boolean('train', False, 'Specify whether to train or evaluate a model.')

Directory Structure for diabetic_retinopathy :

.diabetic_retinopathy
├── ./IDRID_dataset
│   ├── ./IDRID_dataset/CC-BY-4.0.txt
│   ├── ./IDRID_dataset/LICENSE.txt
│   ├── ./IDRID_dataset/images
│   │   ├── ./IDRID_dataset/images/test
│   │   └── ./IDRID_dataset/images/train
│   └── ./IDRID_dataset/labels
│       ├── ./IDRID_dataset/labels/test.csv
│       └── ./IDRID_dataset/labels/train.csv
├── ./Poster_Diabetic_Retionopathy_.pdf
├── ./README.md
├── ./configs
│   └── ./configs/config.gin
├── ./constants.py
├── ./evaluation
│   └── ./evaluation/eval.py
├── ./hp_log_dir
│   ├── ./hp_log_dir/cpts
│   ├── ./hp_log_dir/csv_log
│   ├── ./hp_log_dir/fit
│   ├── ./hp_log_dir/hparam_tuning
│   └── ./hp_log_dir/results
├── ./hyper_parameter_tuning
│   └── ./hyper_parameter_tuning/hparam_tuning.py
├── ./input_pipeline
│   ├── ./input_pipeline/datasets.py
│   ├── ./input_pipeline/datasets2.py
│   └── ./input_pipeline/preprocessing.py
├── ./log_dir
│   ├── ./log_dir/checkpoint
│   ├── ./log_dir/cpts
│   ├── ./log_dir/csv_log
│   ├── ./log_dir/fit
├── ./main.py
├── ./media
├── ./models
│   ├── ./models/architecture.py
│   └── ./models/transfer_learning_architecture.py
├── ./requirements.txt
├── ./results
│   ├── ./results/classification_report.csv
│   └── ./results/confusionmatrix.png
├── ./visualization
│   └── ./visualization/deep_vis.py
└── ./weights
    ├── ./weights/20201222-220802_ADAM_epochs_100_test_acc_78.h5
    ├── ./weights/20201224-212524_ADAM_epochs_10_test_acc_81_55_cm_32_52.h5
    └── ./weights/fullmodel_tl_82_acc.h5