An Improved Light- weight Deep Transfer Learning for Fetal Lung Ultrasound Image Segmentation

Prenatal assessment of fetal lung maturity is challenge and an effective non-invasive way for prenatal evaluation of fetal lung is needed. Ultrasonography has been developed in clinics, but it requires experts, increasing medical cost and doctor's workload. Some machine learning methods, based on U-net, have been studied, but it demands significant computational resources and large-scale dataset. In view of above, this paper proposed a light-weight deep transfer learning architecture, based on improved DeepLabV3+ with MobileNetV2, for fetal lung ultrasound image segmentation. It was trained by 1000pcs of ultrasound image with fetal lung manually delineated and then test on 200pcs ultrasound dataset. By comparison test between U-net, DeepLabV3+ with Xception, DeepLabV3+ with MobileNetV2, we proved our proposed architecture has advanced performance with limited computational resources.