README.md

FundusNet

FundusNet: a deep learning approach for identifying novel endophenotypes for neurodegenerative and eye diseases from fundus images

Hu, W., Li, K., Gagnon, J., Wang, Y., Raney, T., Chen, J., Chen, Y., Okunuki, Y., Chen, W., & Zhang, B. (2025). FundusNet: A Deep-Learning Approach for Fast Diagnosis of Neurodegenerative and Eye Diseases Using Fundus Images. Bioengineering, 12(1), 57. https://doi.org/10.3390/bioengineering12010057

Steps:

1. Install cuda
2. conda create --name <name_of_your_env> or conda create -p <path_of_your_env>
conda activate <name_of_your_env>
   Install pytorch https://pytorch.org/get-started/locally/
   Example, pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu126
3. git clone the repo
   git clone https://github.com/interactivereport/FundusNet.git
4. install packages
   pip install -r requirements.txt
5. Execute either shgender.sh or shage.sh to run individual CNN or ViT models:
   a. This process will split the image dataset into training and testing sets, train the CNN/ViT models on the training data, and evaluate them on the test data.
   b. Users must provide the following inputs:
   'name of csv_file (string)': Path to the CSV file containing annotations.
   'root_dir (string)': Directory containing all images.
6. Combine the results using majority voting for ensemble prediction.

A tutorial for result interpretation using gradCAM:

import os
import numpy as np
import math

import torch
from pytorch_grad_cam import GradCAM 

from modelwrapper import Modelwrapper
from gcam_utils import run_save_gcam_results

# specify input files
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

model_name = 'regnety_32'
pheno_name = 'glaucoma'

# examples of model checkpoints, https://zenodo.org/records/15711816
cmodel_ckpt = f'../model_ckpt/{pheno_name}_{model_name}_v0.pth'
csv_file = f'../result_biomarker_interpretation/input_imgs/imgs_{pheno_name}.csv'
img_dir = '../result_biomarker_interpretation/input_imgs/'
result_dir = '../result_biomarker_interpretation/'

# run gradCAM to generate figures for biomarker interpretation
num_classes = 1 if pheno_name=='age' else 2
mwrapper = Modelwrapper(num_classes=num_classes)
cmodel = getattr(mwrapper, model_name)().to(device)
cmodel.load_state_dict(torch.load(cmodel_ckpt))
cmodel.eval()

target_layers = [cmodel.s4.b1.conv3.conv]

use_cuda = True if device.type=='cuda' else False
cam = GradCAM(model=cmodel, target_layers=target_layers, use_cuda=True)

run_save_gcam_results(cam=cam,
                      img_dir=img_dir,
                      result_dir=result_dir,
                      csv_file=csv_file,
                      pheno_name=pheno_name)

['image_0_glaucoma_1.png', 'image_1_glaucoma_1.png', 'image_2_glaucoma_1.png', 'image_3_glaucoma_1.png']
['image_4_glaucoma_0.png', 'image_5_glaucoma_0.png', 'image_6_glaucoma_0.png', 'image_7_glaucoma_0.png']

A tutorial for predicting age and gender:

import os
import numpy as np
import pandas as pd

import torch

from modelwrapper import Modelwrapper
from imgdataset import ImgDataset_withaugment

# specify input files
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

model_name = 'regnety_32'
pheno_name = 'age'# 'gender'

cmodel_ckpt = f'../model_ckpt/{pheno_name}_{model_name}_v0.pth'
csv_file = f'../result_biomarker_interpretation/input_imgs/imgs_{pheno_name}.csv'
img_dir = '../result_biomarker_interpretation/input_imgs/'
result_dir = '../result_biomarker_interpretation/'

batchsize = 2

num_classes = 1 if pheno_name=='age' else 2
mwrapper = Modelwrapper(num_classes=num_classes)
cmodel = getattr(mwrapper, model_name)().to(device)
cmodel.load_state_dict(torch.load(cmodel_ckpt))
cmodel.eval()

image_datasets = ImgDataset_withaugment(csv_file=csv_file, root_dir=img_dir, crop='center')
dataloader = torch.utils.data.DataLoader(image_datasets, batch_size=batchsize, num_workers=4)
dataloader_iterator = iter(dataloader)

img_names = []
y_true = []
y_preds = []

for inputs, labels, img_name in dataloader:
    img_names.extend(list(img_name))
    y_true.extend(labels.tolist())
    inputs = inputs.to(device)
    labels = labels.to(device)
    preds = cmodel(inputs)
    if num_classes==2:
        _, preds = torch.max(preds, 1)
    y_preds.extend(preds.cpu().detach().numpy().flatten())

df = pd.DataFrame({'filename':img_names, 'class':y_true, f'predicted_{pheno_name}':y_preds})
df.to_csv(os.path.join(result_dir, f'imgs_{pheno_name}_predicted.csv'), index=False)
print(df.head())

             filename  class  predicted_age
0  image_0_age_61.png     61      58.069225
1  image_1_age_61.png     61      62.588211
2  image_2_age_72.png     72      73.754265
3  image_3_age_53.png     53      53.196304