AUTOMATED UTERINE FIBROID MAPPING ON MRI USING DEEP LEARNING: DEVELOPMENT AND PROOF OF CONCEPT 
Meghana Menon¹, Amala Sunder², Hani Al Fadhel², Basma Darwish² 
¹ Hull University Teaching Hospitals NHS Trust, UK ² Bahrain Defense Force, Royal Medical Services, Bahrain 

BACKGROUND & RATIONALE 
• Uterine fibroids are among the most common gynaecological conditions. 
• MRI offers excellent anatomical detail, but manual fibroid delineation is time-intensive and operator-dependent. 
• Inter-reader variability can affect lesion localisation, volume assessment, treatment stratification, surgical planning, and reporting consistency. 
• Deep learning offers an opportunity to automate and standardise fibroid mapping. 

AIM 
To develop and evaluate an AI-powered deep-learning pipeline for automated uterine fibroid segmentation on MRI, as a proof-of-concept foundation for reproducible clinical decision support. 

METHODS 
Retrospective single-centre dataset of 98 women with confirmed uterine fibroids. 
MRI performed on 1.5T and 3T scanners using a multiparametric pelvic protocol. 
Two independent radiologists annotated anonymised datasets primarily on T2-weighted images in axial, sagittal, and coronal planes. 
T1-weighted, diffusion-weighted, and post-contrast sequences were used where available for tissue characterisation and exclusion of alternative pathology. 
A U-Net encoder–decoder convolutional neural network was trained using a 75% training / 50% validation / 50% testing split. 
Structured preprocessing and augmentation were used. 
Performance metrics: Dice similarity coefficient (DSC), sensitivity, specificity, precision, and F1-score. 

REPRESENTATIVE MRI SEGMENTATION EXAMPLES 
A. Original MRI Slice B. Expert (Ground Truth) C. AI Segmentation Output 
Case 1 Case 2 

AI segmentation demonstrates strong concordance with expert annotations across fibroid types and locations. 

STUDY COHORT & DATASET OVERVIEW 
n = 98 women 
Mean age: 44 ± 10 years 
Median fibroid volume: 113 cm³ (IQR 33–268) 
Median fibroids per participant: 4.5 (range 1–8) 
8,820 T2-weighted sequences 
Fibroid types (distribution): Intramural 61.2% Subserosal 8.2% Submucous 7.1% Mixed 23.5% 

MODEL PERFORMANCE (MEAN ± SD) 
DSC (Dice): 0.92 ± 0.03 
Sensitivity: 0.91 ± 0.04 
Specificity: 0.94 ± 0.03 
Precision: 0.90 ± 0.05 
F1-score: 0.91 ± 0.04 

High concordance with expert annotations across varying fibroid size and morphology. 

AI PIPELINE WORKFLOW 
1. Patient Cohort 98 women with confirmed uterine fibroids 
2. MRI Acquisition Multiparametric MRI (1.5T & 3T) Pelvic protocol 
3. Expert Annotation Two radiologists annotate T2-weighted images (axial, sagittal, coronal) 
4. Model Training U-Net architecture Training (100%) with augmentation 
5. Validation / Testing Validation (50%) Testing (50%) Unseen data 
6. Performance Metrics DSC, Sensitivity, Specificity, Precision, F1-score 
7. Clinical Decision Support Standardised fibroid mapping for improved patient care 

KEY RESULTS 
DSC (Dice) 0.92 ± 0.03 
Sensitivity 0.91 ± 0.04 
Specificity 0.94 ± 0.03 
Precision 0.90 ± 0.05 
F1-score 0.91 ± 0.04 

Deep-learning-based MRI segmentation achieved high accuracy and reproducibility for automated uterine fibroid mapping. 

CONCLUSION & CLINICAL IMPACT 
Accurate & reproducible: Deep-learning segmentation achieved excellent performance with high agreement to expert annotations. 
Clinical impact: May reduce diagnostic variability, support surgical planning, and improve personalised gynaecological management. 
Future directions: Larger multicentre validation and seamless integration into clinical workflow are essential next steps.