Sybil

Lung Cancer Risk Prediction

Run the model

You can load our pretrained model trained on the NLST dataset, and score a given DICOM serie as follows:

from sybil import Serie, Sybil

# Load a trained model
model = Sybil("sybil_base")

# Get risk scores
serie = Serie([dicom_path_1, dicom_path_2, ...])
scores = model.predict([serie])

# You can also evaluate by providing labels
serie = Serie([dicom_path_1, dicom_path_2, ...], label=1)
results = model.evaluate([serie])

Models available include: sybil_base and sybil_ensemble.

All model files are available here.

Replicating results

You can replicate the results from our model using our training script:

python train.py

See our documentation for a full description of Sybil's training parameters.

LDCT Orientation

The model expects the input to be an Axial LDCT, where the first frame is of the abdominal region and the last frame is along the clavicles.

When the input is of the dicom type, the frames will be automatically sorted. However, for png inputs, the path of the PNG files must be in the right anatomical order.

Annotations

To help train the model, two fellowship-trained thoracic radiologists jointly annotated suspicious lesions on NLST LDCTs using MD.AI software for all participants who developed cancer within 1 year after an LDCT. Each lesion’s volume was marked with bounding boxes on contiguous thin-cut axial images. The “ground truth” annotations were informed by the imaging appearance and the clinical data provided by the NLST, i.e., the series and image number of cancerous nodules and the anatomical location of biopsy-confirmed lung cancers. For these participants, lesions in the location of subsequently diagnosed cancers were also annotated, even if the precursor lesion lacked imaging features specific for cancer.

Annotations are availble to download in JSON format here. The JSON file is structured as below, where (x,y) refers to the top left corner of the bounding box, and all values are normlized to the image size (512,512).

{
  series1_id: {   # Series Instance UID
    image1_id: [  # SOP Instance UID / file name
      {"x": x_axis_value, "y": y_axis_value, "height": bounding_box_heigh, "width": bounding_box_width}, # bounding box 1
      {"x": x_axis_value, "y": y_axis_value, "height": bounding_box_heigh, "width": bounding_box_width}, # bounding box 2
      ...
      ],
    image2_id: [],
    ...
  }
  series2_id: {},
  ...
}

Cite