Cesarean Scar Defect Segmentation in Transvaginal Ultrasound Images: a Dataset and Benchmark

Technical Summary: Cesarean Scar Defect Segmentation in Transvaginal Ultrasound Images

Problem Statement
Cesarean Scar Defect (CSD), a persistent niche or indentation in the anterior uterine wall resulting from incomplete healing after cesarean delivery, is a prevalent complication affecting approximately 60% of women with a history of cesarean section. While significant defects (residual myometrial thickness <3mm) can lead to symptomatic disease (e.g., postmenstrual spotting, chronic pelvic pain) and serious obstetric risks (e.g., uterine rupture, placenta accreta), accurate diagnosis remains challenging. Transvaginal ultrasonography (TVUS) is the standard first-line screening tool due to its accessibility and non-invasiveness; however, reliable identification is hindered by the small size, irregular morphology, and low contrast of CSD lesions, as well as operator dependency and limited clinical awareness. Despite advances in deep learning for medical imaging, a critical gap exists: there is no publicly available dataset specifically for transvaginal ultrasound CSD segmentation, limiting the development of automated diagnostic tools.

Methodology
To address this data scarcity, the authors present a comprehensive, single-center retrospective cohort study involving the creation of a specialized CSD dataset.

• Data Acquisition: The study utilized anonymized data from the International Peace Maternity and Child Health Hospital (IPMCH). Inclusion criteria focused on women aged 22–45 with significant uterine diverticulum (depth ≥50% of myometrial thickness, residual thickness ≤3mm) and active fertility desires. Raw data was acquired using standardized ultrasound equipment (Philips iU-22, Samsung W10, Mindray Resona R9), yielding 1,111 static images and 16 dynamic video sequences.
• Annotation Protocol: A rigorous, collaborative annotation framework was established involving experienced sonographers and trained PhD students. The process followed standardized clinical guidelines:
  1. Sonographers guided PhD students in identifying CSD presence and regions.
  2. Annotations were performed using Labelme for static images and Pair for video sequences.
  3. A multi-level validation process ensured inter-observer consistency, where each sample was discussed by two PhD students and subsequently reviewed and modified by a sonographer.
• Data Pre-processing: To reduce redundancy, every 10th frame was extracted from the 16 video sequences. A semi-automatic cropping algorithm removed extraneous digital information surrounding the Region of Interest (ROI). From an initial pool of 1,309 samples, 501 high-quality positive samples with confirmed CSD and precise pixel-level manual annotations were selected to constitute the final dataset.
• Dataset Structure: The dataset is organized in the Pascal VOC format, containing original JPEG images, corresponding PNG segmentation masks, and text files defining training/testing splits.

Key Contributions

1. First Public CSD Dataset: The paper introduces the first publicly available dataset for transvaginal ultrasound CSD segmentation, comprising 501 positive samples with pixel-level annotations derived from 1,111 images and 16 videos.
2. Benchmarking Framework: The authors provide a benchmark by evaluating four state-of-the-art medical image segmentation architectures (UNet, DeepLabV3+, GCNet, Swin-UNet) on the dataset using a five-fold cross-validation protocol.
3. Standardized Annotation: The work establishes a reproducible annotation pipeline involving multidisciplinary collaboration, ensuring high-quality ground truth that adheres to clinical guidelines.

Results
The technical validation was conducted using Dice coefficient, Precision, Recall, Intersection-over-Union (IoU), and 95-percentile Hausdorff Distance (HD95).

• Performance Metrics: Across the five folds, the mean Dice coefficients ranged from 74.03% (Swin-UNet) to 75.92% (DeepLabV3+). IoU scores spanned 70.44% to 72.12%.
• Top Performer: DeepLabV3+ achieved the highest performance with a Dice of 75.92 ± 2.99% and an IoU of 72.12 ± 3.06%.
• Error Analysis: Precision varied between 78% and 88%, reflecting the difficulty in distinguishing thin, irregular CSD edges from the background. However, Recall remained robust (minimum fold-mean 78%), indicating that lesion omission was not a dominant failure mode.
• Boundary Accuracy: HD95 distances were consistently below 9 voxels (best result: 6.60 ± 1.90 mm for GCNet), confirming that predicted boundaries remain within a clinically acceptable margin relative to expert traces.
• Stability: Low standard deviations across folds indicate stable model performance and suggest the dataset size is sufficient for robust training.

Significance
The paper posits that this dataset serves as a foundational resource for advancing medical image segmentation algorithms specifically for gynecological ultrasound. By providing high-quality, standardized annotations, the work aims to:

• Enhance the medical community's understanding of CSD characteristics.
• Support the development of automated analysis tools that can reveal quantitative features and subtle patterns potentially missed by human visual detection.
• Improve diagnostic efficiency and reproducibility, ultimately aiding in risk stratification and treatment planning for women of reproductive age.
• Establish CSD segmentation as a novel task within the broader field of medical image analysis, facilitating future research into AI-driven clinical innovations for post-cesarean reproductive health.