Patience
A retrospective analysis of data from Zhejiang People’s Hospital, Hangzhou, Zhejiang, China, was performed and approved by the Zhejiang People’s Hospital Medical Ethics Committee, and all methods were performed in accordance with the principles of the Declaration of Helsinki. In this study, 155 nodules were collected from 148 patients with complex C-SN and ACR BI-RADS category ≥4a in US reports from 2018 to 2021. Data and images were collected in a retrospective study as training group. In addition, this study prospectively collected data and photographs associated with 76 nodules from 72 patients as a test group, according to the training group requirements. The above data were obtained after surgical or percutaneous puncture with corresponding pathological results. Written information was provided and informed consent was obtained from all subjects.
Inclusion and Exclusion Criteria 1. Inclusion Criteria: (a) All breast nodules had a BI-RADS classification of 4a or greater. (b) Ultrasound report listed all nodules as C-SN. (c) All nodes were manipulated to obtain full pathological results. 2. Exclude candidates: (a) A nodule with unclear pathology was found. (b) incomplete or missing clinical and imaging information; (c) Patients underwent radiotherapy, chemotherapy, and fine-needle biopsy prior to US examination. (d) The patient was breast-feeding at the time of US presentation.
After selection, 177 nodules from 170 patients were included in the study. From the Ultrasound Image Archiving and Communication System (PACS) retrospective data, 109 nodes were selected for the training group, including 74 benign and 35 malignant nodes. A total of 68 nodes, including 38 benign and 30 malignant nodes, were selected for the test group from prospective data. Of these, only 59 node images were selected for the DL test group because some images did not meet the requirements. Details are shown in Figure 1. According to the ACR BI RADS classification, 121 (72%) nodes were classified as Grade 4a, 23 nodes (14%) were classified as Grade 4b, and 15 nodes (9%) were classified as Grade 4a. Classified. classified as grade 4c and 9 nodes (5%) were classified as grade 5.
Acquisition of clinical data and ultrasound features
Clinical features include age, height, weight, lactation history, reproductive history, and family history. Among them, age, height, and weight were continuous variables, and lactation history, closure history, and family history were categorical variables. Color ultrasound Doppler images were obtained from multiple ultrasound systems including the Philips Epic 5 ultrasound system (Philips Medical Systems, Bothell, WA, USA), the Supersonic Aixplorer ultrasound system (Supersonic Imagine, Provence, France), Mindray Resona 7, and Mindray. Acquired from the ultrasound machine. DC-8 (Mindray, Shenzhen, China). In this study, we used a high-frequency line array probe with a center frequency of 12 MHz or higher, used only color Doppler modes that require sharp color signals (excluding other blood flow imaging modes such as energy Doppler), and scaler Reduced clutter and color spillover using color. Red/blue scale and 4-8 cm/s scale. Ultrasound features were extracted and judged respectively by her two breast ultrasound specialists, who have more than 5 years of experience in breast ultrasound diagnosis. Features extracted included margins, lesion shape, distribution, aspect ratio, distribution of cystic and solid components, and cystic fluid penetration. , cystic-solid intersection, presence of spongy structures/capsules, microcalcifications, internal vascularity/BF, and the above features were determined and subject to dichotomous variation (0 for negative, 1 for positive). rice field. Disagreements regarding the suitability of trials for inclusion in the review were resolved by consensus by discussion. In this study, the above characteristics of the training group were used as independent variables, and benign and malignant outcomes were used as dependent variables. We created a traditional statistical model using multiple logistic regression, screened independent predictors, tested the accuracy of the model using test groups, and computed the ROC curve.
DL model
In this study, retrospective study data and images from January 2018 to June 2021 were used as a training group (25% randomized data were included as a validation group to guide hyperparameter selection). ), prospectively collected study data and images from July 2021 were used. –August 2022 was used as an independent test group. In this study, we used Resnet50 as a pretrained model.
In this study, the input image size was cropped and normalized to 224*224 pixels, the batch size was 64, and the training cycle was 30 rounds. Training group images were scaled, randomly rotated, randomly cropped, adjusted contrast, adjusted hue, and adjusted saturation using data enhancement mode to reduce the effects of overfitting and sample imbalance. , which significantly increases the number of training samples after data enhancement. There are 1,260 images in the malignant group and 1,404 images in the benign group after augmentation. It continuously updates the model parameters by forward computation and backpropagation to compute the loss function. Validate the training model using images from an independent test group, generate ROC and PR curves, and plot the confusion matrix.
DL that combines clinical and ultrasound functions
In this study, we incorporated two variables derived from the predictions from the traditional statistical model and the predictions from the DL as independent variables into the new logistic regression equation. We calculated the predicted value of CM and plotted the ROC curve. We compared the areas under the ROC curves of these three models to validate the accuracy of the models and to exclude the best models. Figure 6 shows the model building process.
Flowchart of the two models and the process of model combination. FC fully connected layers, ROC receiver operating characteristic curve, PR exact recall, DLRMore Deep Learning Radiomics.
Research on auxiliary functions of CM
In our study, we additionally selected two sonographers with 3 and 5 years of experience in breast ultrasound to identify benign and malignant nodules in the test group. An independent diagnosis was made in the first round, and a re-diagnosis was made in combination with the CM diagnosis in the second round. Diagnostic accuracy with and without CM assistance was also compared.
statistical analysis
In this study, we classified the data into training and test groups, each into benign and malignant groups, and compared baseline data on clinical and ultrasound characteristics. Quantitative data were compared using: t test or mann whitney U, qualitative data were compared using the chi-square test. Consistency test of results judged by two breast ultrasound experts using the Kappa method. Traditional statistical models use multiple logistic regression equations. AUC values are used to compare the diagnostic function performance of the three models. The Hanley & McNeil method was used to compare the diagnostic efficacy of her two sonographers before and after using CM. P. Values less than 0.05 for all statistical data are considered statistically significant. All statistical analyzes were performed using SPSS (SPSS 23.0, SPSS Inc., Chicago, IL), R Studio (R 4.2.1 based), Anaconda 3 (Python 3.9).
