Implementation of our ICIP 2021 paper : CCF-Net: Composite Context Fusion Network with Inter-slice Correlative Fusion for Multi-disease Lesion Detection.
To the best of our knowledge, CCF-Net demonstrates limitations when applied to datasets of moderately large scale. In our recent study, we propose a practical and effective solution to pre-train 3D CNNs with 2D data, which results in highly robust and versatile 3D pre-trained models capable of handling a range of downstream tasks. Further details on our approach can be found here.
This code is based on MMDetection. Please see it for installation.
We perform experiments on the Tianchi Pulmonary Multi-disease datase, which is not publicly avalible now. While, any datasets with coco-style annotations can be used to train a CCF model.
Please run this bash to train a model:
bash ccfnet.sh train