AI ‐Assisted Classification of Phyllodes Tumor from Breast Ultrasound to Reduce Surgical Biopsy
Journal of Ultrasound in Medicine · 2025
85.37% to 87.09%, with sensitivity between 0.72 and 0.91, and specificity between 0.86 and 0.88.When using the consensus of all four readers as the reference standard, the model's accuracy was 86.57%, with sensitivity of 0.95 and specificity of 0.86.Additionally, if images flagged as having artifacts by at least three readers were used as true labels, the model's accuracy was 87.26%, with sensitivity of 0.93 and specificity of 0.87.Conclusions: Our transfer learning model based on the ResNet-18 architecture demonstrated consistent accuracy across different datasets with varying types of transducers and two scanner model for identifying uniformity artifacts.Unlike our previous work, which required separate models for specific transducer, this unified approach simplifies the process and eliminates the need for multiple models.High sensitivity and specificity reported in this study, using independent datasets acquired over several years, highlight the applicability and generalizability of this model to enhance both diagnostic accuracy and efficiency.This approach not only streamlines the workflow but also reduces observer time required for artifact detection.This advancement lays the groundwork for more effective artifact detection in routine practice.