Building a pelvic organ prolapse diagnostic model using vision transformer on multi-sequence MRI

Area of Science:

• Pelvic floor imaging analysis
• Deep learning in medical diagnostics
• Female pelvic medicine

Background:

• Pelvic organs like the uterus, bladder, and rectum share interconnected fascia, leading to potential shared risk factors and associations for prolapse.
• Organ prolapse can impact neighboring pelvic organs, necessitating accurate assessment of disease severity.
• Current manual measurements for prolapse severity are subjective and can lead to diagnostic inaccuracies.

Purpose of the Study:

• To develop a deep learning-based multilabel grading model for classifying pelvic organ prolapse (POP).
• To utilize stress magnetic resonance imaging (MRI) for assessing prolapse in three pelvic organs.
• To provide interpretable results for the developed grading model.

Main Methods:

• Utilized sagittal MRI sequences from 662 subjects during rest and Valsalva maneuver.
• Developed a vision transformer-based feature extraction module for pelvic floor MRI.
• Employed label masking and pre-training strategies for model convergence, evaluating with Precision, Kappa, Recall, and AUC.

Main Results:

• The model achieved high performance metrics: average Precision of 0.86, Kappa of 0.77, Recall of 0.76, and AUC of 0.86.
• Outperformed existing studies in POP grading detection.
• Demonstrated a rapid average diagnosis time of 0.38 seconds per patient.

Conclusions:

• The proposed deep learning model demonstrates accuracy comparable to or exceeding physicians in grading POP.
• The vision transformer architecture and label masking strategy are effective for POP grading under static and Valsalva conditions.
• This AI model presents a promising tool for computer-aided diagnosis and treatment planning of pelvic organ prolapse.