This repository contains a TensorFlow implementation of the 169-layers Densely Connected Convolutional Networks (DenseNet-169) model for the task of abnormality detection in musculoskeletal Radiographs on MURA dataset.
The weights of the model are initialized from a pretrained model on ImageNet using caffe, DenseNet-Caffe. The caffemodel weights are then extracted as Numpy arrays and saved as a pickle file using caffe_weight_converter. The extracted weights can be downloaded here, or generated using the above-mentioned repositories.
The network is fine-tuned end-to-end using Adam optimiser with default parameters (β1 = 0.9 and β2 = 0.999), and an initial learning rate of 1e-4 which is divided by 10 every time the validation loss plateaus. Two training strategies have been tested. First, the network is trained on the entire dataset, resulting in a single model which is used for inference. Second, the network is trained on each of the seven body parts (elbow, finger, forearm, hand, humerus, shoulder, wrist) in the dataset, resulting in seven part-specific models which are ensembled for inference.
- Training on the entire dataset :
<python train.py --bpart=all>
- Training on the "elbow" dataset:
<python train.py --bpart=elbow>
- Single model evaluation:
<python evaluate.py --bpart=all>
- Ensemble model evaluation:
<python evaluate_ensemble.py --bpart=all>
-
Validation set
-
Single model
Body part AUROC Accuracy Kappa All 0.8625 0.8040 0.6024 -
Ensemble model
Body part AUROC Accuracy Kappa All 0.8797 0.8307 0.6556
-
-
Test set
- Single model
- Ensemble model
@article{rajpurkar2018mura, title={MURA Dataset: Towards Radiologist-Level Abnormality Detection in Musculoskeletal Radiographs}, author={Rajpurkar, Pranav and Irvin, Jeremy and Bagul, Aarti and Ding, Daisy and Duan, Tony and Mehta, Hershel and Yang, Brandon and Zhu, Kaylie and Laird, Dillon and Ball, Robyn L and others}, year={2018} }
@inproceedings{huang2017densely, title={Densely connected convolutional networks}, author={Huang, Gao and Liu, Zhuang and van der Maaten, Laurens and Weinberger, Kilian Q }, booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition}, year={2017} }