A mutual promotion encoder-decoder method for ultrasonic hydronephrosis diagnosis

As a common cause of hydronephrosis in children, ureteropelvic junction obstruction (UPJO) may lead to a series of progressive renal dysfunction. Ultrasonography is a primary screening of UPJO, yet its further examinations are laborious, time-consuming, and mostly radioactive. The deep learning based automatic diagnosis algorithms on UPJO or hydronephrosis ultrasound images are still rare and performance remains unsatisfactory owning to limitation of manually identified region of interest, small dataset and labels from single institution. To relieve the burden of children, parents, and doctors, and avoid wasting every bit information in all datasets, we hence designed a deep learning based mutual promotion model for the auto diagnosis of UPJO. This model consists of a semantic segmentation section and a classification section, they shared a mutual usage of a transformation structure by separately training the encoder and decoder and loop this circle. Thorough comparative experiments are conducted and situations are explored by ablation experiments, results shown our methods outperformed classic networks with an accuracy of 0.891 and an F1-score of 0.895. Our design can jointly utilize different supervisions and maximize the use of all the characteristics of each dataset, and automatically diagnose the severity of UPJO on the basis of ultrasound images by first segmentate then classify the images, moreover, not only is the final result excellent, but also the midway segmentation result is also very accurate and have smooth edges that are convenient for doctors to recognize with their naked eyes. All in all, our proposed method can be an important auxiliary tool for smart healthcare.