| dc.description.abstract | Computer-aided diagnosis, such as radiomics, includes methods that are primarily based on pattern recognition, which increase the diagnostic accuracy of the physician and improve the computer diagnosis obtained from the quantitative analysis of radiological images. The aim of this study is to evaluation of diagnostic value of radiomix MRI images of bone lesions in differentiating benign from malignant lesions.
Materials and Methods: The present study was cross-sectional and used a database of MRI images of patients referred with benign and malignant bone lesions (with pathological confirmation) of any age and gender during the years 2016 to 2022 to one of the related teaching hospitals in Tabriz, including Tabriz Children's Hospital, Tabriz Imam Reza Hospital, or Tabriz Ghazi Hospital. The minimum sample size was determined to be 150 people. The sampling method in this study was available. MRI images should have been taken before surgery or neoadjuvant treatment in all patients. Demographic information as well as the exact diagnosis, tumor type and side involved in the disease, tumor stage, and metastases were extracted from the patient records. Then, the patients' DICOM files were imported into the 3D Slicer software, and masking, segmentation, and determination of tumor boundaries were performed independently by two radiologists. If there was a difference between the two radiologists, the average of the obtained values was taken. Finally, the radiomics calculated by 3D Slicer software was entered into SPSS statistical software for analysis and comparison between the two diseases.
Results: In this study, 150 patients in two groups of 75 with benign and malignant bone lesions were studied, and the mean (standard deviation) age in the two groups of patients with benign and malignant bone lesions studied was 45.06 (±18.6) and 43.12 (±9.6) years, respectively. In this study, among the 9 radiomics models considered to differentiate benign from malignant lesions in the patients studied, the best validation accuracy overall and in patients over 45 and under 45 years of age in all three was related to the VGG16 model with 73.2%, 71.3%, and 75.3%, respectively, and the best AUC of all three was related to the InceptionV3 model with 54.1%, 54.4%, and 57.3%, respectively. Based on the VGG16 model, which has the best validation accuracy, overall and in patients over 45 and under 45 years of age, the sensitivity was 79.1%, 80.2%, and 81.4%, the specificity was 70.2%, 70.3%, and 72.2%, the positive predictive value was 51.5%, 52.5%, and 50.6%, and the negative predictive value was 54.4%, 55.5%, and 55.6%, respectively. | en_US |
