Investigations Of Medical Image Classification Based On Partial Features And Attention Mechanism

Smart medical systems have shown great advantage in improving efficiency,accuracy and safety for medical treatments.In the current years,there is a broad prospective to reduce the cost reduction and improve medical services efficiency.Medical image assisted diagnosis technology is one of the promoting applications of artificial intelligence in medical treatments.As an important technical of assisted diagnosis method,the information obtained from medical images needs clinicians in a long time clinical experience.Traditional image processing software cannot obtain valuable pathological information from images,but can only realize functionalities,such as image reading,writing,and length measurement based on pixels.As a result,enormous manual analysis is necessary to obtain essential information,which obviously increases the effort of clinicians.However,the technologies based on artificial intelligence can realize functionalities,such as symptom recognition,lesion localization,and severity estimation.Moreover,using intelligence algorithms can also reduce subjective influences and effectively improve the diagnosis speed as well as the efficiency of medical treatment.Traditional medical image analysis techniques mainly rely on machine learning algorithms,which require manual design of features.The effectiveness of these features depends on the designer’s prior and professional knowledge.When the distribution of input data has a slightly departure with respect to the prior,previous features are invalid,resulting in reduced generalization capability.On the contrary,the image processing algorithms based on deep learning can autonomously extract effective semantic features from images,avoiding manually designed features.Thus,the accuracy and generalization ability of the model are further improved.Inspired by these advantages,in this thesis,we use the deep learning methodology to study medical image classification.The residual network 50(Res Net50)is taken as the baseline model,and three different methods are proposed to improve the discrimination capacity of the model.Intensive experiments are done based on Chest X-ray dataset(CXI dataset)and APTOS2019-Blindness-Detection dataset(ABD dataset).First,a medical image classification algorithm based on local feature is proposed.The main idea of the algorithm is to extract semantic features of images by dividing the last feature tensor horizontally into several parts and locally pooling each part into a vector.The loss of each part is calculated separately,and the total loss is calculated by summing up these losses to optimize the whole network,so as to improve the discrimination ability of extracting local features.In experiments,Res Net50,Res Net50+PP6(dividing the last feature tensor into 6 parts)and Res Net50+PP8(dividing the last feature tensor into 8 parts)are trained on CXI dataset and ABD dataset respectively.Compared with the baseline model,the present algorithm improves the classification accuracy to 8.3% and 1.5%respectively.The experimental result indicates that the medical image classification algorithm based on local features can improve the classification ability of the model to a certain extent.Second,from the knowledge of attention mechanism,a novel spatial attention mechanism based on inner product is proposed.This structure can be embedded into the medical image classification model.The conventional spatial attention mechanism uses maximum pooling and average pooling to refine the feature tensor along the channel direction.This leads to the loss of local subtle features and reduces the feature extraction and recognition capacity of the model.The present attention mechanism utilizes the inner product between the column feature vectors along the channel direction and a learnable vector to preserve local subtle information.These subtle information is encoded in the output feature tensor along the channel direction,thus reducing the feature loss,which may be helpful to improve the feature extraction ability of the network.In experiments,Res Net50 with conventional spatial attention and Res Net50 with the proposed inner product attention models are trained on CXI and ABD datasets respectively.Compared with the baseline model,the present algorithm improves the classification accuracy to 1.3% and3.9% respectively.Finally,inspired by fine-grained image classification methods,the idea of applying the sliding window and non-maximum suppression to medical image classification tasks is proposed.In this way,a medical image classification model based on attention position proposal mechanism(APPM)is presented.In this method,only image-level labels are used without pixel-level labeling information.Moreover,the conventional sliding window method is combined with convolutional neural network to find local windows with richest semantic information.The feature is extracted separately from these windows and fused with the global feature from the whole image.In experiments,a model of Res Net50 with APPM is trained on CXI dataset and ABD dataset respectively.It is found that the proposed algorithm improves the accuracy on CXI and ABD datasets by 13.3% and 8.6%,respectively,compared with the baseline model.In addition,the corresponding confusion matrix and ROC curve are calculated for each model.These experimental results combined with the dimensional reduction visualization results of the extracted features of the models indicate that the three methods proposed in this thesis not only can extract more effective features from medical images,but also improve the overall performance of the model.