Bone-Abnormality-Classification (Homepage)
Bone abnormalities classification is crucial in diagnosing Musculoskeletal Disorders (MSDs). In this research, Regnet is used with a three-layer classifier to predict the abnormality of the hand X-ray image. Data augmentation such as rotation, horizontal flip, and translation are proved to benefit the model in this task. To deal with the data imbalance, we also propose a weighted binary cross-entropy loss function. Learning rate is decay in step to achieve the local minimum of the model. Grad-Cam is used to visualize the abnormality of specific region in the image. Overall, we reach an AUC:0.82 in testing data on Kaggle after ensembling.
please download regnet.pt model from github and put in the same folder as train.py and inference.py
train.py:
Inference.py :
Example : python "/data1/home/8B07/Anthony/bone-abnormality-classification/final/inference.py" --data /data1/home/8B07/Anthony/bone-abnormality-classification/final/ -output /data1/home/8B07/Anthony/bone-abnormality-classification/final/test.csv
requirements.txt is provided in the folder
General network structure in the RegNet
The X block in the RegNet model, which consists of the residual bottleneck block and group convolution
Grad-Cam of abnormal image. Obviously, the abnormality of this image is the internal fixation device, which corresponding to the yellow spot on the heat map
Grad-Cam of fingertip amputations. The model mainly focus on the marks rather than amputation site