What is this?
This is the official PyTorch code for the MICCAI 2023 STS Challenge (3D tooth segmentation from CBCT scans) by team sdkxd, which extends nn-UNet with axial attention and a positional correction module to clean up abnormal segmentation points.
This is the official PyTorch code for the MICCAI 2023 STS Challenge (3D tooth segmentation from CBCT scans) by team sdkxd, which extends nn-UNet with axial attention and a positional correction module to clean up abnormal segmentation points.Is this a standalone tool or does it extend nn-UNet?It is not a standalone application: it's a set of modified files and scripts you copy into an existing nn-UNet installation (e.g. a patched genericUNet.py) to add the attention and correction modules.
It is not a standalone application: it's a set of modified files and scripts you copy into an existing nn-UNet installation (e.g. a patched genericUNet.py) to add the attention and correction modules.How do I get the pipeline running?First install nn-UNet and organize your labeled CBCT data into its required Task folder structure (dataset.json, imagesTr, imagesTs, labelsTr), run nn-UNet's automatic preprocessing, then train with 5-fold cross validation before moving on to pseudo-labeling and nnUNet-att training.
First install nn-UNet and organize your labeled CBCT data into its required Task folder structure (dataset.json, imagesTr, imagesTs, labelsTr), run nn-UNet's automatic preprocessing, then train with 5-fold cross validation before moving on to pseudo-labeling and nnUNet-att training.Do I need nn-UNet installed before using this?Yes, this code is built entirely on top of nn-UNet and assumes you already have it installed and meeting its requirements; no extra dependencies beyond that are needed.
Yes, this code is built entirely on top of nn-UNet and assumes you already have it installed and meeting its requirements; no extra dependencies beyond that are needed.Is this a one-command pipeline?No, it's a multi-stage manual workflow: train on labeled data, generate pseudo labels for unlabeled data, iteratively retrain and filter low-quality pseudo labels, then train nnUNet-att and finally run the position correction step on the predictions.
No, it's a multi-stage manual workflow: train on labeled data, generate pseudo labels for unlabeled data, iteratively retrain and filter low-quality pseudo labels, then train nnUNet-att and finally run the position correction step on the predictions.How well did this solution perform in the challenge?On the final unseen test set, team sdkxd's solution ranked 5th overall with a score of 81.47%.
On the final unseen test set, team sdkxd's solution ranked 5th overall with a score of 81.47%.What does the positional correction module do?It's a post-processing step that detects and fixes abnormal segmentation points in the predicted tooth masks after inference.
It's a post-processing step that detects and fixes abnormal segmentation points in the predicted tooth masks after inference.Why bother generating pseudo labels?The challenge has far more unlabeled CBCT scans than labeled ones, so the pipeline trains on labeled data first, then iteratively generates and filters pseudo labels on unlabeled scans to expand the training set semi-supervised style.
The challenge has far more unlabeled CBCT scans than labeled ones, so the pipeline trains on labeled data first, then iteratively generates and filters pseudo labels on unlabeled scans to expand the training set semi-supervised style.What is nnUNet-att?It's the nn-UNet architecture modified with axial attention layers (via a patched genericUNet.py) and trained with 5-fold cross validation on the combined labeled plus filtered pseudo-labeled data.
It's the nn-UNet architecture modified with axial attention layers (via a patched genericUNet.py) and trained with 5-fold cross validation on the combined labeled plus filtered pseudo-labeled data.What are similar projects?VisionHOPE, Attention-Residuals, pointdit, External-Attention-pytorch, awesome-semantic-segmentation, Engram, deep-learning-for-image-processing, CVPR2026-Papers-with-Code
VisionHOPE, Attention-Residuals, pointdit, External-Attention-pytorch, awesome-semantic-segmentation, Engram, deep-learning-for-image-processing, CVPR2026-Papers-with-Code