Automatic Ultrasonic Image Classification For Small Sample And Unbalanced Data

As one of the most commonly used screening diagnostic tools in clinical practice,ultrasound has been recognized as the preferred imaging method for breast cancer and thyroid cancer.With the development of deep learning technology,the application of computer vision and pattern recognition technology to computer-aided diagnosis and analysis of clinical ultrasound has become a research hotspot in the field of medical imaging.However,there are problems such as poor quality of ultrasound data imaging,unclear edge details,large amount of speckle noise,small data problem and small scales,and significant differences in the order of magnitude between positive and negative samples in data sets.Automatic classification is challenging.So automatic classification by computer-aided diagnosis is challenging.In order to solve the small sample problem of ultrasound image dataset,this thesis explores transfering ultrasonic image datasets to the medical ultrasound,based on the natural image field of AlexNet,VGG-16 and the SphereFace model based on the face image domain.The classification accuracy on the thyroid ultrasound dataset were 89.3%,91.33% and 93.54%,respectively.In addition,this thesis proposes sample augumentation based on GAN network.At first,the data is augmented by traditional methods.The augmented data set is input into GAN for feature learning,and high-quality ultrasound image samples are generated as a new augmented data set.The experimental results show that the augmented data set generated by GAN sample gains the best result in the benign and malignant classification experiments.The classification accuracy is 14.29%,which is higher than the original unaugumented ultrasound data set,the sensitivity is improved by 17.94%,and the specificity is improved by 12.87%.The accuracy increased by 3.57%,the sensitivity increased by 2.56%,and the specificity increased by 3.96% compared with the traditional data enhancement method.It proves that the use of GAN for sample generation can enrich the diversity of data,thus improve the classification effect of the data set.On the other hand,for the problem of the imbalance of benign and malignant categories in the ultrasound image data,this thesis uses SMOTE algorithm to oversample the minority samples.After this operation,the accuracy rate on the breast ultrasound data set is increased by 2.82%,and the sensitivity is improved by 10.61%.The specificity was increased by 0.46%;the accuracy of the thyroid ultrasound data was increased by 3.14%,the sensitivity was increased by 1.72%,and the specificity was increased by 10%.In order to improve the single ultrasound data classification effect of sample categories,this thesis proposes an SVM classification method combining deep features.Compared with the traditional neural network classification method,the accuracy of breast ultrasound data is improved by 0.14%,which is accurate on thyroid ultrasound data,the accuracy has increased by 0.86%.Compared with the algorithm using the same thyroid public ultrasound dataset,the accuracy of this algorithm is 4.71% and the sensitivity is improved by 8.02%,which proves the effectiveness of the proposed algorithm in solving small sample and class imbalance problems.