pelvitect-seg — Pediatric Pelvic Segmentation (SegFormer-B3)

Part of the Pelvitect framework for automated analysis of developmental dysplasia of the hip (DDH) from anteroposterior pelvic radiographs.

Model description

SegFormer-B3 fine-tuned on PelviSet (10.5281/zenodo.20615290), a unified dataset of 14,622 pediatric pelvic radiographs with 12-class anatomical segmentation masks. The model predicts pixel-level anatomical structures used downstream by the Pelvitect geometry engine to compute clinical measurements (acetabular index, center-edge angle, Tönnis grade, IHDI grade, Shenton line).

Label schema (13 classes including background)

Index	Structure
0	Background
1	Right ilium
3	Right obturator foramen
5	Right pubis
7	Right ischium
9	Right femoral head
11	Right proximal femur

Recommended usage

# Via Pelvitect (recommended — handles full pipeline)
from pelvitect import Pelvitect
import numpy as np
from PIL import Image

img = np.array(Image.open("xray.png").convert("RGB"))
result = Pelvitect(image=img).analyze()
print(result.summary())

# Mask only (no keypoints)
from pelvitect.analyzer import SegmentationMasker
masker = SegmentationMasker()          # downloads this repo automatically
mask   = masker.predict_mask(img)      # (H, W) uint8, values 0-12
lm, shenton = masker.derive_all(mask)  # geometry landmarks + Shenton arcs

Training details

• Base model: nvidia/segformer-b3-finetuned-ade-512-512
• Dataset: PelviSet — MTDDH DS1+DS2 (real masks) + USTC-DDH + AV-DDH (pseudo-labeled via iterative nnUNet v2 pipeline with Bridge-IoU gates)
• Real mask training cases: 1,857
• Expanded corpus: 13,980 cases (real + pseudo-labeled)
• Horizontal flip disabled: laterality-aware training (flip would swap right/left labels, causing training collapse)
• Input size: 512 × 512 (resized by SegformerImageProcessor)

Citation

@dataset{pelviset2025,
  author    = {Mahdavikia, Amir M.},
  title     = {PelviSet: A Unified Pediatric Pelvic Radiograph Dataset},
  year      = {2025},
  doi       = {10.5281/zenodo.20615290},
}