The Research On Classification Of Ultrasound Breast Tumor Based On Particle Swarm Optimized And Multi-layer Perceptron

Breast cancer is one of the most common malignancies among women in the world,which develops from breast tissues.On account of its high mortality,early screening and diagnosis is particularly important.By virtue of its convenience,low cost,no radiation trauma and other advantages,ultrasonography has become an important means of early screening for breast cancer.It can greatly avoid unnecessary biopsy.However,the reading of breast ultrasound images heavily depends on the experience of radiologists’.It also takes a great deal of effort and resources to train experienced radiologists,so the computer-aided diagnostic(CAD)system come into existence.Nevertheless,traditional CAD systems are all based on low-level imaging features,like texture and morphologic features.It may result in a sharp decline in the performance of the diagnostic system when the source of ultrasound images change.Furthermore,the given classification results are difficult to be understood and accepted by physicians.Therefore,this paper proposes the human-in-the-loop CAD model to overcome the aforementioned disadvantages of traditional methods.However,this method relies heavily on manual scoring.For the same breast tumor ultrasound image,sonographers with different experiences may give different scores,which may affect the final classification results.To a certain extent,the above problems will limit the clinical application and promotion of ultrasonography in the early screening and diagnosis of breast cancer.Aiming at the huge semantic gap between the low-level imaging features used in traditional diagnostic systems and the high-level semantic features understood by physicians,this thesis proposes a diagnostic system which based on BI-RADS feature scoring scheme combining the experience of experts.By utilizing Biclustering algorithm and fuzzy inference,the diagnostic process of radiologists can be simulated.Later,the system parameters were optimized by using particle swarm optimization algorithm.Finally,we can establish a benign and malignant classification system of breast tumors which has strong interpretability and high reliability.In view of the problem that the cognitive differences among sonographers at different experience levels may affect the classification results,a cleaning method of breast tumor scoring data that based on causal inference and multi-layer perceptron was proposed.It can reduce the misclassification effect caused by wrong scores.Subsequently,random forests was conducted to classify breast cancer.In this dissertation,1488 cases of ultrasound breast tumor were used to verify the performance of proposed methods.The experimental results show that both of the two schemes have better classification performance and can provide valueable help for sonographers in the clinical diagnosis.