treblenalto/bone-fracture-classification

Identifying bone fracture using deep convolutional neural networks

Bone Fracture Detection

Overview

Binary classification of bone CT images between normal/fractured bone using covolutional neural networks (CNN). A Total of 353 images (177 fractured, 176 normal) were used for training, with 20% as validation. All images were center-cropped to 400 x 400 dimension. Due to the small size of data, training images were augmented using random horizontal flip(p=0.5).
Variants of VGGNet, AlexNet, and ResNet were trained to classify the images. Out of the three, AlexNet achieved best validation performance, thus was selected as final model to predict test images.

This was carried out as a course project for "Pattern Recognition and Machine Learning" taught by Prof. Taehoon Shin at Ewha Womans University.

Data

Train Data

• Fractured Images: CT images of fractured bones

• Normal Images: CT images of normal bones

Test Data

• Mix of fractured and normal bone images

Model

• VGGNet - Very Deep Convolutional Networks For Large-Scale Image Recognition, 2014, (Karen Simonyan, Andrew Zisserman)
• AlexNet - ImageNet Classification with Deep Convolutional Neural Networks, 2012, (Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton)
• ResNet - Deep Residual Learning for Image Recognition, 2015, (Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun)

Final Model

Modified AlexNet with input size 400x400, Dropout(p=0.8) trained with L2 Regularization(lambda=1e-3)

All of the models were trained and tested via Pytorch framework.
No pretrained model was used for this project.

Results

Accuracy	Loss

Structure

Bone Fracure Detection
├── README.md
├── code
│   ├───AlexNet.ipynb
│   ├───AlexNet_Drop.ipynb
│   ├───AlexNet_Drop_L2Reg(1e-3).ipynb
│   ├───AlexNet_Drop_L2Reg(1e-4).ipynb
│   ├───ResNet16.ipynb
│   ├───ResNet26.ipynb
│   ├───VGG9.ipynb
│   └───VGG11.ipynb
│
├── report.pdf
│