Han Wu, Haoyuan Chen, Lin Zhou, Qi Xu, Zhiming Cui, Dinggang Shen
IEEE Transactions on Medical Imaging, 2026
DOI: 10.1109/TMI.2026.3651389

SemiEchoTracker trains with a full echocardiography sequence as input while using landmark supervision only on the first and last frames. The model contains:

• a spatial detector that directly regresses landmark coordinates for every frame;
• a bidirectional temporal tracker initialized from labelled endpoint landmarks;
• endpoint supervision for both detector and tracker;
• temporal tracker regularization with forward/backward consistency and velocity consistency;
• detector-tracker co-training on unlabeled middle frames;
• frozen guided-DINOv2 features with perception-aware spatial-temporal attention.

At inference time, only the detector branch is used to predict landmarks over the whole sequence.

dataset/echo_dataset.py
models/
  SemiEchoTracker.py
  GCN.py
  agent_attention.py
  dinov2.py
train.py
test.py

The guided-DINOv2 pretraining procedure is not included; models/dinov2.py expects pretrained DINOv2 weights to be available.

Set SEMI_ECHO_PRETRAIN_DIR to the pretrained DINOv2 directory:

export SEMI_ECHO_PRETRAIN_DIR=/path/to/dinov2-finetune

The loader expects the hub files and checkpoint in that directory:

$SEMI_ECHO_PRETRAIN_DIR/
  hubconf.py
  teacher_checkpoint.pth
  ...

Semi-supervised training:

python train.py \
  --data_dir ../data/ \
  --data_type PLAX \
  --training_mode semi

Detector-only baseline:

python train.py \
  --data_dir ../data/ \
  --data_type PLAX \
  --training_mode detector_only

test.py evaluates detector predictions. The checkpoint can be either a run directory or a .pth file.

python test.py \
  --checkpoint ./results/PLAX/YOUR_RUN \
  --data_dir ../data/ \
  --data_type PLAX \
  --split val

@article{wu2026semiechotracker,
  title={Semi-Supervised Landmark Tracking in Echocardiography Video via Spatial-Temporal Co-Training and Perception-Aware Attention},
  author={Wu, Han and Chen, Haoyuan and Zhou, Lin and Xu, Qi and Cui, Zhiming and Shen, Dinggang},
  journal={IEEE Transactions on Medical Imaging},
  year={2026},
  doi={10.1109/TMI.2026.3651389}
}