Transfer Learning for 3-Dimensional Medical Image Analysis

The accuracy of deep learning models is significantly affected by the size of training data available. Models pre‐trained from massive dataset speed up training convergence and betters the performance of the model. Similarly, tools based on high volume datasets are necessary for the development of deep learning in machine learning in 3‐Dimensional medical images. Nonetheless, it is very difficult to build a large enough dataset because of the difficulty of data acquisition and annotation in 3‐Dimensional medical imaging. We accumulate the dataset from considerable challenges to build 3D Seg‐8 dataset with varied modalities, targeted organs, and pathologies. To extract threedimension (3‐Dimensional) features, we design a heterogeneous 3‐Dimensional mesh (or network) to co‐train our dataset so as to make a series of pre‐trained classifiers. Experiments show that our model can further the training convergence twice as compared with model pre‐trained on Kinetics dataset, and about tentuple in comparison with training from scratch as well as attain accuracy in the range of 3 to 20%. Transferring our model on a state‐the‐of‐art DenseASPP segmentation network, in case of a single model, we achieve 94.6%.