Rapid classification of fetal ultrasound images based on deep learning, decision fusion and quantization

Background and Aims: Classification of fetal ultrasound images is important for prenatal care and diagnosis, but manual classification is time-consuming and depends on operator experience. While previous studies have shown good results using deep learning algorithms, existing models often require large computational resources, which limits their implementation in clinical settings. The aim of this study was to develop and evaluate an automated fetal ultrasound image classification system using transformer-based models and quantization that requires low computational resources while maintaining high accuracy and is suitable for real-world implementation. Methods: In this study, a classification algorithm was developed by combining the Vision Transformer (ViT) and Swin Transformer models, which is optimized for ۸-bit quantization-aware training (QAT). The model was trained and analyzed using a public dataset with ۱۲,۴۰۰ images of fetal ultrasounds in six groups (abdomen, brain, femur, chest, maternal cervix, others). The performance of the model was evaluated using top-۱ and top-۳ errors. In addition, generalizability of the model was tested using a local dataset consisting of ۹۲ fetal ultrasound images collected from one of the ultrasound centers in Arak city. Results: The proposed model achieved an accuracy rate of ۹۳.۶% on the test dataset, which was consistent with the performance of previous models. The proposed model was significantly improved in terms of computational efficiency. The model developed in this study performed ۵ times better than DenseNet-۱۶۹ and ۲.۵ times better than Inception-v۳ and ResNeXt-۱۰۱. Furthermore, in the local dataset, the model achieved an accuracy of ۶۷.۳۹%, indicating acceptable generalization to new clinical settings. Conclusion: The use of transformer-based models together with quantization can achieve adequate classification quality while at the same time significantly reducing the