Cost-Efficient Light-Weight YOLO V5_s for Whole Fetus Detection in Early-Stage Ultrasound Scans

Abstract

Routine medical Ultrasound (US) scans are recommended for expectant mothers to monitor the health and growth of the fetus. However, expectant mothers in rural areas of developing and underdeveloped countries face difficulties in receiving timely scans due to a lack of expertise and facilities. Consequently, maternal and fetal deaths occur at higher rates in these countries, especially in the first trimester. Novel concepts such as Virtual Doctors, Hospital to Home, self- scans and the Internet of Medical Things (IoMT) may address the aforementioned problem effectively. Therefore,
computational-efficient algorithms which support low-end smart devices should be introduced to assist expectant mothers in rural areas to provide comfortable and timely fetal scans. In light of the above, this paper discusses computationally efficient YOLO V5_s for fetal detection in first-trimester ultrasound images using a highly diverse dataset including abnormal and multiple pregnancies. The implemented model was compared with five benchmark detection models, namely, ResNet-50 and MobileNet-based faster R-CNN, YOLO-n, YOLO-m and YOLO-l. YOLO was comparatively better than faster R-CNN. Even though YOLO-n is the most computationally inexpensive model, its mAP is 0.709, which is comparatively low, hence cannot be applied to the clinical set-up. YOLO-l model has the best performance with F-1 score and mAP of 0.978 and 0.751, respectively. However, YOLO-s has also achieved a F-1 score of 0.979 with a mAP 0.734. Therefore, a subjective test was conducted to verify using YOLO-s in the clinical set-up with five experts in the field with more than five years of experience. The subjective analysis test, assessed through Fleiss Kappa, suggests substantial agreement beyond chance (κ = 0.69), while the Intraclass Correlation Coefficient (ICC) indicates modest reliability (ICC = 0.7). The findings endorse the application of YOLO-s for real-time detection of whole fetuses in the first trimester with reduced computational complexity with further validation.