Abstract
BackgroundTraditional dynamic contrast-enhanced MRI (DCE-MRI) radiomics approaches for predicting lymphovascular invasion (LVI) in invasive breast cancer (IBC) frequently neglect the importance of dynamic phase alterations, and their diagnostic efficacy is often constrained by limited sample sizes. We have developed the Tabular Prior-data Fitted Network (TabPFN) algorithm, which synergistically combines clinical and MR morphological features with delta-radiomics, thereby substantially improving the performance of binary classification.MethodIn this retrospective study, 276 IBC patients were divided into a training group (n = 193, 70%) and a validation set (n = 83, 30%). A radiomic score (Radscore) was developed using 1239 radiomic features derived from lesion masks in delta images, establishing the delta-radiomics model. To preoperatively predict LVI, we utilized the TabPFN algorithm alongside traditional machine learning methods. This approach combined the Radscore with both clinical and MR morphological features for binary classification.ResultsThe delta-radiomics model achieved an area under the curve (AUC) of 0.775. Among the evaluated machine learning models, the TabPFN algorithm demonstrated superior performance by effectively integrating the Radscore along with clinical and MR morphological features, resulting in an AUC of 0.899. Additionally, it recorded an accuracy of 0.88, a precision of 0.667, a recall of 0.571, and an F1-score of 0.615.ConclusionDelta-radiomics analysis shows potential for predicting preoperative LVI in IBC patients. To tackle small sample sizes, we developed the TabPFN algorithm, combining clinical and MR morphological features with Radscore, enhancing binary classification and demonstrating strong predictive performance.