Research On Breast Cancer Recognition Method Based On Small Sample Learning

Medical image processing,as a long-term focus of academia and industry,has high research and application value.The introduction of computer-aided diagnosis system in medical clinical practice can effectively reduce the pressure of doctors and improve the accuracy of clinical diagnosis.With the rapid development of artificial intelligence technology,medical image processing based on neural network is more and more widely used in clinical diagnosis.Most of the computer-aided diagnosis systems based on neural network rely on massive data which is for training and learning,in order to achieve satisfactory results of diagnosis and discrimination.Some diseases are difficult to obtain a large number of effective data to train the learning algorithm,such as those with low incidence rate,or expect to get effective control of outbreaks at the early stage,such as SARS,Ebola,or COVID-19.In the reality of insufficient samples,the learning algorithm based on massive data training cannot meet the needs of the actual situation.Therefore,as an important branch of neural network which focuses on solving the problem of insufficient samples,small sample-oriented learning method has been paid more and more attention.Due to the small amount of data in the medical small sample application environment,and the difficulty of obtaining it,this thesis takes the breast ultrasound image as the research object,simulates the application scene of small sample by controlling the number of known label samples,and studies the tumor segmentation and classification in the medical image aided diagnosis system for small sample problem.It can not only meet the conditions of small medical samples in actual applications,but also make full use of sufficient test data to adequate verify the method.The main research results are as follows:1.Tumor segmentation method of breast ultrasound image based on minimum barrier distanceIn the tumor segmentation task,the idea of visual saliency detection is re-introduced,and the human visual prior knowledge is used to locate and segment the target tumor,so as to avoid the algorithm failure caused by lack of experience or samples.At the same time,the texture removal algorithm is used as the preprocessing step of ultrasound image,which provides a good prerequisite for the later stage of tumor segmentation.2.Small sample breast ultrasound image classification learning method based on contrastThrough the idea of comparison,we can further achieve the ideal balance of the number of categories while expanding the number,and solve the problem of insufficient number of samples and extreme imbalance of sample categories caused by small samples.This thesis proposes a model independent small sample learning and training process,which replaces the task target of prediction label in traditional deep learning with the consistency of prediction label,and achieves the classification target by predicting the label consistency between the test image of unknown label and the contrast image of known label.3.Medical image small sample learning method based on multi size contrast fusionIn this thesis,a small sample learning method for medical images based on multi-scale contrast fusion is proposed.At the same time,the convolution neural network is used to extract the image features of the contrast image in different sizes for fusion,and the label consistency is also used as the output result to promote the difference between the network learning class samples and reduce the demand for the number of samples.This thesis introduces the idea of visual saliency test and contrast,and proposes a method for breast ultrasound tumor segmentation and classification under the condition of small samples.Experimental results show that the results of this thesis can achieve better segmentation and classification results.Finally,the novel coronavirus pneumonia CT image and melanoma image were used to verify the portability of the classification method based on small sample learning.