Radiomics analysis based on semi-automatic image segmentation of ultrasound for preoperative evaluation of Mammotome-assisted minimally invasive resection

Abstract

Abstract Previous radiomics studies still relied on manual delineation. DeepLabv3_resnet50 and FCN_resnet50 are deep neural networks commonly used for semantic segmentation in recent years. This study evaluated the effects of two different networks for semi-automatic segmentation of ultrasound images, and established an ensemble model. Purpose Performing a preoperative assessment prior to Mammotome-assisted minimally invasive resection to aid physicians in guiding more precise individualized treatments, our research aims to develop an ultrasound-based semi-automatic segmentation ensemble learning model for preoperative assessment Methods From November 2018 to November 2023, we retrospectively collected preoperative ultrasound images from 733 patients and randomly assigned them to two cohorts in an 8:2 ratio: a training cohort and a testing cohort. Among these, 230 patients without breast tumors were also randomly divided into training and testing cohorts following the same 8:2 ratio. We then selected DeepLabv3_resnet50 and FCN_resnet50 models for semi-automatic image segmentation. Radiomic features and deep transfer learning features were extracted from both semi-automatic segmentation outcomes to construct radiomic models, deep learning models, and deep learning radiomic models. An ensemble learning strategy was employed to integrate the deep learning radiomic models from both pathways with clinical models. The predictive performance was evaluated using receiver operating characteristic curves and decision curve analysis. Results The semi-automatic segmentation model, DeepLabv3_resnet50, achieved a peak global accuracy of 99.4% and an average Dice coefficient of 92.0% at its best epoch. On the other hand, the FCN_resnet50 model exhibited a peak global accuracy of 99.5% and an average Dice coefficient of 93.7% at its best epoch.In the task of predicting tumor and non-tumor patients, the stacking model ultimately demonstrated an AUC of 0.890 in the training cohort (with a sensitivity of 0.844 and a specificity of 0.815) and an AUC of 0.780 in the testing cohort (with a sensitivity of 0.713 and a specificity of 0.739).In the task of predicting adenosis and other lesion types, the stacking model achieved an AUC of 0.890 in the training cohort (with a sensitivity of 0.613 and a specificity of 0.859) and an AUC of 0.771 in the testing cohort (with a sensitivity of 0.759 and a specificity of 0.765). Conclusion Our study has established an ensemble learning model grounded in semi-automatic segmentation techniques. This model accurately distinguishes between tumor and non-tumor patients preoperatively, as well as discriminates adenosis from other lesion types among the non-tumor cohort, thus providing valuable insights for individualized treatment planning.