A Dual-Task Deep-Learning Model with Fused Ultrasound Images for Simultaneous Typing and Grading of Cystocele.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Urogynecology

Background:

• Cystocele diagnosis and staging are crucial for effective treatment.
• Current diagnostic methods may lack comprehensive accuracy.
• Automated diagnostic tools can improve efficiency and consistency.

Purpose of the Study:

• To develop and evaluate FD-Net, a dual-task deep-learning model for simultaneous cystocele typing and grading.
• To assess the diagnostic performance of FD-Net using fused 2D and 3D ultrasound images.
• To compare FD-Net's performance against single-modal deep-learning models.

Main Methods:

• A retrospective study included 625 patients (467 cystocele, 158 normal).
• FD-Net fused preprocessed 2D (resting and Valsalva) and 3D (levator hiatus) ultrasound images.
• The model performed typing (normal, type I/II/III) and grading (normal, mild, significant) tasks, compared to single-modal models (ST-Net, SG-Net).

Main Results:

• FD-Net demonstrated higher accuracy in typing (79.68%) and grading (81.38%) compared to 2D-only models.
• Significant improvements in F1-scores were observed for normal and mild cystocele cases.
• Area under the ROC curve (AUC) values exceeded 0.92 for all categories.

Conclusions:

• The fused dual-task FD-Net model achieves superior diagnostic performance for cystocele typing and grading.
• FD-Net shows significant potential for clinical application in urogynecology.
• Image fusion in deep learning enhances diagnostic accuracy for complex medical conditions.