Quality evaluation of ultrasound images for fetal crown-rump length measurement at 11 to 14 weeks' gestation: A fusion model can be interpreted using SHAP method.

Abstract

Objectives: To explore a fusion model designed for the quality evaluation of ultrasound images utilized in fetal crown-rump length (CRL) measurement, and to use SHapley Additive exPlanations (SHAP) method to elucidate the model's decision-making processes.

Methods: We retrospectively collected 1149 images of midsagittal planes of the entire fetus during early pregnancy from two hospitals. Two senior radiologists categorized the images into standard and non-standard planes. Seven image segmentation models were trained to select the best model for automatically segmenting the region of interest. The radiomics features and deep transfer learning (DTL) features were extracted and selected to establish radiomics models and DTL models. We also constructed fusion models to enhance the classification performance and the optimal one underwent comparison with radiologists. The SHAP method was employed to interpret and visualize the model.

Results: The DeepLabV3 ResNet101 segmentation model demonstrated the highest performance (DSC: 97.15%). The early fusion model exhibited superior classification performance in validation set (AUC: 0.947, 95% CI: 0.924-0.970, accuracy: 88.4%, sensitivity: 83.0%, specificity: 92.7%, PPV: 90.1%, NPV: 87.3%, precision: 90.1%). The model demonstrated performance commensurate with that of senior radiologists while surpassing junior radiologists. Notably, when leveraging the model's support, there was a substantial improvement in their overall performance.

Conclusions: The early fusion model demonstrated satisfactory performance in the intelligent quality evaluation of ultrasound images for CRL measurement. It has the potential to enhance the professional skills of junior radiologists.