Musculoskeletal Radiographs Abnormality Detection

Bone X-Ray Deep Learning Competition. Read more here.

This was my first attempt at this challenge, I used Xception architecture for this and received a Cappa score of 0.687 on the test set. Leaderboard Model name: xception(single model) bimal

Metrics based on study type per image in the Validation Set

Xception model received an accracy of 83% on the Validation set.

===== ELBOW ======
roc_auc_score:  0.8655411655874191
Sensitivity:  0.9702127659574468
Specificity:  0.7608695652173914
Cohen-Cappa-Score:  0.7327090673094752
F1 Score:  0.8803088803088802
===== FINGER ======
roc_auc_score:  0.7782454879110069
Sensitivity:  0.8925233644859814
Specificity:  0.6639676113360324
Cohen-Cappa-Score:  0.546259842519685
F1 Score:  0.7827868852459016
===== FOREARM ======
roc_auc_score:  0.8475496688741722
Sensitivity:  0.96
Specificity:  0.7350993377483444
Cohen-Cappa-Score:  0.6945780209114572
F1 Score:  0.8622754491017965
===== HAND ======
roc_auc_score:  0.7469005642437376
Sensitivity:  0.959409594095941
Specificity:  0.5343915343915344
Cohen-Cappa-Score:  0.5262369439474013
F1 Score:  0.840064620355412
===== HUMERUS ======
roc_auc_score:  0.8754826254826255
Sensitivity:  0.8581081081081081
Specificity:  0.8928571428571429
Cohen-Cappa-Score:  0.7500964134207482
F1 Score:  0.8758620689655172
===== SHOULDER ======
roc_auc_score:  0.7810299129117759
Sensitivity:  0.8210526315789474
Specificity:  0.7410071942446043
Cohen-Cappa-Score:  0.5625817210210153
F1 Score:  0.7918781725888325
===== WRIST ======
roc_auc_score:  0.8414648910411623
Sensitivity:  0.9642857142857143
Specificity:  0.7186440677966102
Cohen-Cappa-Score:  0.6986424046110609
F1 Score:  0.8796992481203008

Steps to train the Neural Network: (Train this on the GPU)

1. Clone this project.

2. Download the dataset from the competetion home page and place it in the cloned project directory. For example: Musculoskeletal-Radiographs-abnormality-detection/

3. Run the main_train.ipynb.

Project Dependencies

1. pip install tensorflow-gpu
2. pip install keras
3. pip install scipy
4. pip install sklearn
5. pip install pandas
6. pip install numpy
7. pip install jupyter

Neural Net was trained in AWS using the p2.xlarge instance, I have created an AMI which comes with all of the above dependencies installed. AMI-ID: ami-0337b3fba4a212c7f, make sure you’ve selected the Ireland region if you decide to use this AMI.

Project Structure:

1. src/data/dataloader.py: Filters out normal and abnormal x-ray images and adds them to their own directories for keras ImageDataGenerator.flow_from_directory(). For example train/normal/ and train/abnormal/

2. src/data/postprocessor.py: Avergaes out the probablities of images per patient study type.

3. src/model/*: Creates a Xception or Dense169 neural network model

4. src/predict.py: Run this when ready for submission. Note: Change the value of the trained_model_path variable.