A mutual promotion encoder-decoder method for ultrasonic hydronephrosis diagnosis.

Ureteropelvic junction obstruction (UPJO) is a frequent cause of hydronephrosis in children, potentially leading to progressive kidney dysfunction. While ultrasonography is widely used for initial screening, further diagnostic steps remain labor-intensive and often involve radiation. Traditional methods rely on manual identification of regions of interest, which is time-consuming and limited by small datasets and single-institution labeling. These constraints hinder the development of robust automatic diagnosis systems. Prior research has shown that manual segmentation and classification are error-prone and inefficient. No prior work had resolved the issue of maximizing data utility in UPJO diagnosis. This gap motivated the exploration of deep learning-based solutions. However, existing models lack performance due to limited training data and poor generalization. The need for a more efficient and accurate diagnostic tool remains unmet. This paper addresses the challenge of improving UPJO detection through a novel deep learning framework.

The aim of this study is to develop a deep learning model for automatic diagnosis of UPJO using ultrasound images. The researchers propose a mutual promotion encoder-decoder architecture to enhance diagnostic accuracy. The specific problem addressed is the inefficiency and inaccuracy of manual diagnosis methods. The motivation stems from the need to reduce diagnostic burden on doctors and parents while improving diagnostic precision. The study focuses on maximizing the use of available data and improving segmentation accuracy. The researchers aim to create a system that can be used in smart healthcare settings. Their approach seeks to improve both classification and segmentation performance simultaneously. The ultimate goal is to provide a reliable diagnostic tool for UPJO in children.

The researchers designed a mutual promotion encoder-decoder model for UPJO diagnosis. The model includes a semantic segmentation module and a classification module. These components share a transformation structure, allowing mutual information exchange. The encoder and decoder are trained separately in a looped process. Ablation experiments were conducted to evaluate model performance. Comparative experiments were performed against classic networks. The model uses ultrasound images as input for diagnosis. The design allows for joint utilization of different supervision signals.

The proposed model achieved an accuracy of 0.891 and an F1-score of 0.895. These results outperformed classic networks in UPJO diagnosis. The model's segmentation results showed smooth edges and high accuracy. The classification results were also excellent, indicating strong diagnostic performance. The mutual promotion mechanism improved both segmentation and classification. The model maximized the use of available data characteristics. The segmentation results were useful for visual recognition by doctors. The model demonstrated high potential as a diagnostic aid.

The authors propose that their mutual promotion model improves UPJO diagnosis accuracy. The model's design allows for efficient use of available data. The segmentation results are accurate and visually interpretable. The classification results are reliable and clinically useful. The model's performance outperforms traditional methods. The mutual promotion mechanism enhances diagnostic accuracy. The model is suitable for smart healthcare applications. The study suggests the model can serve as an important diagnostic aid.