| Breast cancer(BC)is one of the most common malignant tumors in women,with a heavy burden and a high mortality rate.In the decision-making support of diagnosis and treatment for BC,breast ultrasound is a non-invasive,low-cost,real-time,portable,radiationless,and internal tissue structure imaging examination.Ultrasound images contain histopathologic information at tissue and even molecular levels.However,the information is usually difficult to be directly observed visually.Radiomics analysis at the intersection of medicine and engineering integrates artificial intelligence(AI)technologies such as computer vision and machine learning,which can quantitatively and objectively dig into the potential knowledge related to histopathologic characteristics connotative in ultrasound images.This can assist in accurate screening,intelligent diagnosis,and personalized treatment decisions for BC,becoming a potentially effective means to prevent and reduce the incidence and improve patients’ prognostic quality.This dissertation sequentially analyzes the specific clinical issues of BC diagnosis and treatment decision support with ultrasound images-based AI technology for early screening and identification,surgical treatment and metastasis,and survival prognosis,and proposes a few novel approaches to address the remaining difficulties and deficiencies of these issues.These novel approaches have achieved superior results and mainly include the following.1.Aiming at the problem that developing AI tools to distinguish ductal carcinoma in situ(DCIS)and fibroadenoma(FA)based on routine ultrasound images requires high-cost annotation,a novel difference-based self-supervised(DSS)learning approach is proposed.The DSS approach only requires FA samples that are common and easy to label in routine ultrasound examinations to participate in training.It builds an anomaly detection network on a self-supervised network to reconstruct an image by encoding and decoding,and detect the size of differences between the reconstructed and original images to classfy images.Experimental results show that compared with conventional AI models,the DSS-based model 1)achieved better receiver operating characteristic curve(Receiver Operating Characteristic,ROC)and area under the curve(Area Under ROC curve,AUC),2)its calibration curve expressed an acceptable level of deviation between the predicted and observed values,and 3)its decision curve could provide patients with higher net benefits across a wider range of risk thresholds.The DSS approach can reduce the high-cost annotation work without sacrificing diagnostic performance,thus promoting the development of ultrasound-based AI screening tools to identify DCIS and helping to advance BC disease control to an earlier stage.2.Aiming at the problem that preoperative prediction of ipsilateral supraclavicular lymph nodes(ISLN)metastasis using ultrasound-based radiomics analysis is a data mining problem with a small sample size and an ultra-high-dimensional feature space,a novel approach termed GALambda is proposed to search for a linear feature combination.The GALambda approach utilizes a genetic algorithm to search feature subspaces,and then performs high penalty factor regularized LASSO regression from the feature subspaces and select linear feature combinations by dynamically adjusting the penalty factor(i.e.,lambda)to select linear combinations of features.The feature combinations serve as stable genes for individual crossover in the genetic algorithm to optimize and iterate populations,and finally search for a feature combination that meets a preset object.Experimental results show that compared with conventional feature selection approaches,the GALambda-based model 1)achieved higher ROC&AUC,2)its calibration curve expresses a basic consistency between the predicted and observed values,indicating that the model may have strong robustness and generalization ability,and 3)its decision curve could provide patients with higher net benefits across a wider range of risk thresholds.The GALambda approach is rewarding for the radiomics analysis from the dataset with a small sample size and an ultra-high-dimensional feature space,and can be used to support the prediction of ISLN metastasis in BC patients based on preoperative ultrasound.This approach is also a reference for similar radiomics studies.3.Aiming at the problem that preoperative prediction of 5-year disease-free survival(DFS)for triple-negative BC(TNBC)patients using ultrasound-based radiomics analysis is a data mining problem with a small sample size and an extremely imbalanced class distribution,a novel approach termed contourlet coefficient count(CCCount)is proposed to extract features.The details of local changes in ultrasound images often reflect structural information of pathological tissue.Considering the sample imbalance,CCCount fuses the anisotropic characteristics of contourlet coefficients at each decomposition level based on single-class training samples,then count the fused coefficients.Then,it uses the counting results as baselines to discretize the fused coefficients and finally extracts quantitative features with statistical methods to characterize local change information.Experimental results show that compared with conventional feature extraction approaches,the CCCount-based model 1)achieved higher ROC&AUC,its calibration curve expresses a basic consistency between the predicted and observed values,and the deviation degree is the smallest among all models,indicating that the CCCount-based model can behave with stronger robustness and generalization capabilities,and 3)its decision curve could provide patients with higher net benefits across a wider range of risk thresholds.The CCCount approach is rewarding for the radiomics analysis from the dataset with a small sample size and an extremely imbalanced class distribution,and can be used to support the prediction of 5-year DFS in TNBC patients based on preoperative ultrasound.This approach is also a reference for similar radiomics studies. |
PDF Full Text Request