Research On Several Key Techniques Of Breast Tumor Diagnosis Based On Mammograms

Breast cancer is the most common malignancy among women."Early found and early treated" is the most effective way to prevent breast cancer at present,and mammary gland molybdenum target X-ray examination is the main method of early breast cancer check.Doctors diagnose the existence of early breast cancer by reading the mammograms.However,the result of such early diagnosis is mainly based on the subjective opinions of doctors,which may cause a high diagnosis error rate and a high lesion missing rate.Therefore,it is of great importance to improve the efficiency and accuracy of the diagnosis from mammograms for doctors.Along with the development of the computer science technology,computer aided detection and diagnosis technology comes into being.It combines the medical image technology and computer science technology together,and provides auxiliary opinions for doctors to make the diagnosis conclusions,helping to improve the efficiency and accuracy of breast cancer diagnosis.This paper further researches on the computer aided detection and diagnosis technology from the following three subfields with the knowledge of machine learning,pattern recognition technology and image process.(1)Mammogram classification.Classifying the mammograms in image level directly can be achieved with a few annotated mammograms and some mammograms without annotations,considering the fact of the scarcity of annotated mammograms.This paper proposes a mammogram classification method based on the fusing features.The method extracts the global features and the local information of the mammograms and fuses them together,and finally classifies the mammograms with the fusing features.The fusing features expresses the image from a more comprehensive view,which is the reason why it achieves a better classification performance comparing to the global features and the local information.(2)Detection of the breast masses.Different breast masses usually have different sizes,and effective methods of breast mass detection should be able to detect all these masses in different sizes.This paper proposes a breast mass detection method base on multi-scale features.The method utilizes SSD which has shown great performance in natural image detection,and combines the detection results of different feature layers to get the final detection result.Target predictions of different feature layers aim at targets in different sizes of the original image,and it is effective for the detection of masses in different sizes as a result.(3)Segmentation of the breast masses.The main difference between benign masses and malignant masses is the margin.Using mass segmentation to get the specific margin of masses,is an important step of mammogram analysis.This paper proposes a mass segmentation method based on a fully convolutional multi-layer model.The method combines multi-layer features together to segment the mass region,and utilizes Conditional Random Fields to optimize the margin.Low-level features represent more detail information,and high-level features represent more abstract and more global information.Combining them together achieves a more accurate result of the mass segmentation.