Deep Learning Method Incorporating Prior Knowledge And Its Applications

In recent years,deep learning methods have achieved great success in many fields,but such methods completely abandon knowledge and are very ineffective when fac-ing complicated problems such as image blind deblurring.For this reason,we combine prior knowledge with deep learning method,and carry out research on topics such as predefined convolutional filter network,multi-instance learning method based on graph convolutional network and application of new method in medical image analysis.These new studies have achieved good results,open up a new idea of deep learning method research by integrating knowledge,and provide a basic support for medical image anal-ysis and related applications.Convolutional neural network(CNN)is a common framework that is widely used in many practical application tasks.However,to deal with many real-world problems,especially medical image analysis problems,it is necessary to add some task-specific domain knowledge.In this paper,we propose a new convolutional neural network based on domain knowledge:predefined convolutional filter network(PCFNet),in which convolutional kernels of the first convolutional layer are replaced by parameterized trainable image filter kernels.We find that PCFNet can achieve the same approxi-mation accuracy as ordinary convolutional neural networks from a theoretical analy-sis viewpoint when learnable parameters are fewer.Experimental results show that PCFNet achieves high accuracy on the CIFAR10/100 dataset,surpassing many convo-lutional neural network frameworks,which indicates the effectiveness of new network.Moreover,PCFNet achieves state-of-the-art results on handwriting recognition dataset(USPS)and medical image dataset(IDRiD),which shows that for processing some special task,such as handwriting recognition and medical image classification,when choosing appropriate image filter,PCFNet is superior to ordinary convolutional neural network.In addition,experimental results on fracture detection show that two-stage method based on Faster R-CNN and Schmid convolutional filter network achieves high recall and F1 score,surpassing other two-stage methods,which reflects that our method is highly sensitive to fracture line,thus achieving a more accurate fracture recognition rate.Multi-instance learning(MIL)is a method that learns the mapping from instance bags to bag labels,and relationship among instances is an important factor in learning this mapping.In this paper,we combine structural knowledge contained in the data to propose a new multi-instance learning method:multi-instance learning based on graph convolutional network(GCN-based MIL).We first use structural relationship among instances to establish graph structure in bag,and then use graph convolutional network and graph-attention mechanism to learn bag-embedding.Based on the fundamental the-orem of continuous function on the permutation invariance,we theoretically prove that GCN-based MIL has arrangement invariance.Experimental results show that GCN-based MIL achieves high accuracy on five multi-instance learning benchmark datasets,surpassing other multi-instance learning methods,which reflects the effectiveness of our method.In addition,GCN-based MIL has achieved very high accuracy and AUC on four medical image datasets,surpassing other related methods,which shows that our method is more suitable for processing medical high-resolution image classification.In addition,experimental results on melanoma detection and lesion region segmentation(ISIC 2017 tasks)show that GCN-based MIL obtains 0.93 AUC and 0.699 JA,respec-tively,surpassing other related methods.It shows that our method can effectively assist melanoma diagnosis.