NPUBXY solution of the LLD-MMRI Challenge

Preprocessing

• Begin by restoring the original spacing for each NII file using the script misc/remake_original_image.py.
• Conduct a visual inspection to ensure the accuracy of all images. Utilize the code in misc/visual_image.py. I have found some cases in the training set are flipped upside down. These misoriented images were subsequently corrected manually.
• Perform image registration on all eight-phase images. The reference image for registration is the C+pre phase, and all other phases are aligned to this reference. I use CrossSAM (code link) for registration.
• The registered patches, along with our trained model, have been uploaded via Baidu Disk (pw:299z).

Solution

• Our approach utilizes a similar framework as the baseline. However, we replace the 3D UniFormer with a 2D ResNet18 architecture. During training, we randomly select 3 consecutive slices from each cropped case, the output features for all 8 phases are fused by the final FC layer.
• For inference, we only use the center 3 slices of each case.
• To enhance our model's performance, we employ a two-level model ensemble strategy. For a model trained using a specified train/validation split, we identify the top five models exhibiting the highest overall F1-score and kappa values (referred to as best_fk_checkpoint). Additionally, we train models using distinct train/validation splits. In our final solution, we incorporate a total of seven different train/validation splits.

Data and pre-trained models

• The data_model_labels.zip (Baidu Disk pw:299z) contains three folders:

data_model_labels
  - images_mycrop_8 (contains all cropped registered patches with margin of 8)
  - labels (training and validation data split)
  - models (all pretrained models)

Installation

The installation is same as the baseline.

torch==1.13.1
torchvision==0.14.1
SimpleITK==2.0.0rc2.dev910+ga138e
torchio==0.18.91

Inference

• Our training regimen involves the utilization of seven distinct train/validation split models. For each of these models, we select the top five checkpoints based on their F1 and kappa scores. Consequently, a total of 35 models need to be utilized for inference. The final result is obtained by calculating the average of predictions from each individual model.
• The predictions for each model have been uploaded to the models folder, organized as follows:

models
  - resnet18m-1
    - NPUBXY_fk_0.json
    - NPUBXY_fk_1.json
    ...

• To reproduce our results, you can conveniently execute the script python misc/emsem_result_subtraining.py (remember to change the path). This will automatically generate the final average prediction using all NPUBXY_fk_*.json files across all seven train/val split models. The output of this process will precisely match the results submitted during the testing phase.
• Alternatively, you have the option to use the script multi_model_predictions_subtraining.py to generate predictions for each individual model. It's important to note that we employ test-time random augmentation, which might lead to slight variations in the results.

Training

• For training new models with our designated train/validation split, please make use of the code multi_model_training.py.