MedMNIST Active Learning Model

Overview

This model is designed for image classification tasks within the medical imaging domain, specifically targeting the MedMNIST dataset. It employs a ResNet-50 architecture tailored for 28x28 pixel images and incorporates active learning strategies to enhance performance with limited labeled data.

Model Architecture

Base Model: ResNet-50
Modifications:
- Adjusted initial convolution layer to accommodate 28x28 input images.
- Removed max pooling layer to preserve spatial dimensions.
- Customized fully connected layer to output predictions for 9 classes.

Training Procedure

Training Hyperparameters

Hyperparameter	Value
Batch Size	53
Initial Labeled Size	3559
Learning Rate	0.01332344940133225
MC Dropout Passes	6
Samples to Label	4430
Weight Decay	0.00021921795989143406

Optimizer Settings

The optimizer used during training was Stochastic Gradient Descent(SDG), with the following settings and a Learning Rate Scheduler of ReduceLROnPlateau:

learning_rate = 0.01332344940133225
momentum = 0.9
weight_decay = 0.00021921795989143406

The model was trained with float32 precision.

Dataset

PathMNIST

Data Augmentation

Random resized cropping
Horizontal flipping
Random rotations
Color jittering
Gaussian blur
RandAugment

Active Learning Strategy

The active learning process was based on a mixed sampling strategy:

Uncertainty Sampling: Monte Carlo (MC) dropout was used to estimate uncertainty.
Diversity Sampling: K-means clustering was employed to ensure diverse samples.

Evaluation

The model was evaluated on the validation set of PathMNIST. Key performance metrics include:

Accuracy: 94.72%
Loss: 0.2397
AUC: 99.73%

Graphs

The following plots illustrates the validation loss, validation accuracy, and validation auc over batches(number of iterations over the dataset) during the active learning process.

Validation Loss
Validation Accuracy
Validation AUC

Usage

All code for this model can be accessed in the following GitHub Repository: Allen Cheung Determined_AI_Hackathon

To utilize this model:

Install Dependencies: Ensure the following Python packages are installed:
- torch
- torchvision
- medmnist
- scikit-learn
- determined
Install them using pip:
```
pip install torch torchvision medmnist scikit-learn determined
```

Load the Model:

import torch
from model import ResNet50_28

model = ResNet50_28(num_classes=9)
model.load_state_dict(torch.load('pytorch_model.bin'))
model.eval()

Inference:

from torchvision import transforms
from PIL import Image

transform = transforms.Compose([
    transforms.Resize((28, 28)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5], std=[0.5])
])

image = Image.open('path_to_image.jpg')
input_tensor = transform(image).unsqueeze(0)
output = model(input_tensor)
prediction = output.argmax(dim=1).item()
print(f"Predicted class: {prediction}")

License

This project is licensed under the MIT License.

Acknowledgements

MedMNIST Dataset
Determined AI
Survey on Deep Active Learning: Wang, H., Jin, Q., Li, S., Liu, S., Wang, M., & Song, Z. (2024). A comprehensive survey on deep active learning in medical image analysis. Medical Image Analysis, 95, 103201. https://doi.org/10.1016/j.media.2024.103201