The kidney tumor segmentation model is trained on contrast CT (corticomedullary and nephrogenic phases) to accurately delineate the kidney, tumor, and cysts. The KiTS 2023 dataset is employed to train a kidney tumor segmentation model (N=489) using an ensemble of fivefold cross-validation within the nnUNet framework. In the first stage, the trained model generated annotations for 64 cases from TCGA-KIRC collection. These annotations were refined by non-experts. Among these, 45 cases added to the KiTS 2023 dataset as the training set for the second stage model. The final model is used to generate annotations for 156 cases from TCGA-KIRC collection which includes 43 cases from first stage (27 train and 16 test TCGA-KIRC cases). Additionally, radiologists annotated the remaining 39 out of the 156 cases, enabling a comparison between the final model's segmentation and radiologist annotations.TCGA-KIRC collection.
The model_performance notebook contains the code to evaluate the model performance on the TCGA-KIRC collection against a validation evaluated by a radiologist and a non-expert.
As an extended project, this model was used to gennerate annotations an addition IDC Collections. The performance on these collections can be found in the model_performance_aimi2 notebook.
The model was released as a docker container on MHub.ai
you can run the model with the following command, where $in is the path to the input directory of dicom files for the CT scan, and $out is the path to the output directory where the DICOM-SEG will be placed.
docker run --rm -t --gpus all -v $in:/app/data/input_data -v $out:/app/data/output_data mhubai/bamf_nnunet_ct_kidneyDownload an example CT scan from the TCGA-KIRC using idc-index
from idc_index import index
from pathlib import Path
Path("example_data").mkdir(exist_ok=True)
client = index.IDCClient()
client.download_from_selection(
seriesInstanceUID=["1.3.6.1.4.1.14519.5.2.1.2932.1975.291646103899971840049683623517"],
downloadDir="example_data",
)Then run the docker container with the following command
mkdir example_output
export in=$(pwd)/example_data
export out=$(pwd)/example_output
docker run --rm -t --gpus all -v $in:/app/data/input_data -v $out:/app/data/output_data mhubai/bamf_nnunet_ct_kidney