Developing a transfer learning model for the prediction of Down syndrome from first-trimester ultrasound images

Down syndrome is the most common chromosomal disorder, frequently associated with increased nuchal translucency (NT) and an absent nasal bone (NB) during first-trimester ultrasound. This study aims to develop a transfer learning (TL) model for the automated detection of fetuses at high risk for Down syndrome.

An artificial intelligence (AI)-based classification framework was developed using TL with the AlexNet architecture and integrated with gradient-weighted class activation mapping (Grad-CAM) to automatically detect NT and NB features in first-trimester ultrasound images. The model was trained on a dataset comprising 1056 images of normal fetuses and 12 images of fetuses with Down syndrome. To enhance training consistency and model generalizability, image cropping and class weighting techniques were applied.

To address class imbalance, the 12 Down syndrome images were augmented using rotation, mirroring, and brightness adjustments, resulting in 96 anatomically faithful images. The model achieved excellent performance, with an accuracy of 98.3 %. The area under the receiver operating characteristic curve (AUC) was 0.98, and the sensitivity, specificity, positive predictive value, and negative predictive value were 89.47 %, 100 %, 100 %, and 97.96 %, respectively. Grad-CAM visualization indicated that the NT and NB regions were the most influential features in the model's decision-making, providing valuable insights into its diagnostic reasoning.

This TL model enhances the early detection of Down syndrome from first-trimester ultrasound images, improving diagnostic accuracy and reducing the reliance on manual measurements.