Automated Detection and Characterization of Kidney Cysts with Deep Learning from Ultrasonographic Images

Abstract

Ultrasonography is frequently used to identify disorders affecting internal organs since it is non-invasive, non-radioactive, real-time, and reasonably priced. To measure organs and tumours in ultrasonography, a set of measurement markers is positioned at two different places. The size and location of the target finding are then determined using this information. Renal cysts, which affect 20–50% of patients regardless of age, are one of the measurement objectives of abdominal ultrasonography. Because kidney cysts are frequently measured in ultrasound imaging, automating the measurement would also have a significant influence. Creating a deep learning algorithm that can quickly detect kidney cysts in ultrasound pictures and forecast the positions of two crucial anatomical markers to gauge the cysts' sizes was the aim of this study. The deep learning method used refined UNet++ to forecast saliency maps, which show the positions of important landmarks, and refined YOLOv5 to detect kidney cysts. After identifying the ultrasound images, YOLOv5 clipped them into the bounding box and sent them to UNet++. Three sonographers visually positioned important landmarks on 100 test data items that were not visible in order to compare the outcomes with human performance. A board-certified radiologist said that these well-known, famous locations provided the ground truth. Next, we assessed and contrasted the deep learning model's and the sonographers' accuracy. Measurement error and precision-recall metrics were used to assess their performances. The analysis's findings demonstrate that our deep learning model's accuracy and recall for identifying renal cysts are on par with those of conventional radiologists; the radiologists' accuracy and speed in predicting the crucial landmark placements were on par with ours.