Tools Catalogue
      Back to home
      1. Knowledge Base
      2. Image Analysis Management
      3. Tools Catalogue

      Stroke Decision Support Tool (CASE)

      The  Stroke Decision Support Tool was developed as part of the Cloud CT Image Analysis in Stroke Emergency (CASE) grant

      This tool predicts the outcome of the patient with intra-arterial treatment and without intra-arterial treatment 90 days after stroke. The output of the workflow will help to determine the benefit of endovascular treatment.

      The tool expects one of the following combinations of input data:

      • non-contrast CT Image + age
      • Age + NIHSS(baseline) + pre-mRS
      • non-contrast CT Image + age + NIHSS(baseline) + pre-mRS
      • Age + NIHSS(baseline) + pre-mRS + ASPECT
      • non-contrast CT Image + age + NIHSS(baseline) + pre-mRS + ASPECT
      • Age + NIHSS(baseline) + pre-mRS + ASPECT + NIHSS(24h)
      • non-contrast CT Image + age + NIHSS(baseline) + pre-mRS + ASPECT + NIHSS(24h)
      • Age + NIHSS(baseline) + pre-mRS + ASPECT + NIHSS(24h) + NIHSS(1w)
      • non-contrast CT Image + age + NIHSS(baseline) + pre-mRS + ASPECT + NIHSS(24h) + NIHSS(1w)

      NOTE: All the numerical input data is extracted from the metadata fields of the patient.

      The tool outputs a categorized Modified Rankin Score (mRS). The outcome of the patient is categorized as:

      • Good outcome: corresponding to a mRS 90 days after the stroke below or equal to 2.
      • Bad outcome: corresponding to a mRS 90 days after the stroke above 2

      The confidence for each prediction is presented in the form of a Colum Chart.

      Different Machine Learning models with different combinations of data, as well as different hyperparameters, have been studied using the MRClean dataset. For each of the data combinations, we selected the model that performs best in terms of F1-score. When non-contrast CT is available, high-level image features are extracted using a Deep Convolutional Autoencoder. The Autoencoder was trained using non-contrast CT images from the MRClean dataset. 

      The workflow includes a pre-processing pipeline for the non-contrast CT images when they are available for the prediction. The pre-processing pipeline reorients the image to standard orientation, enhances the contrast, aligns the image to the OASIS-30 template, extracts the brain on the CT scan expressed in HU units, normalizes the image to 0-1 range and masks the CT image using the OASIS-30 template priors 4. The workflow also includes DICOM files to NIfTI conversion.

      The prediction of the outcome will be computed using two Machine Learning models depending on the available input data and whether if the model is considering intra-arterial treatment or not.

      Example of the report when using baseline metadata:

      Required inputs

      Minimum:

      • Age
      • NIHSS(baseline)
      • pre-mRS

      Optional: 

      • ASPECT
      • NIHSS(24h)
      • NIHSS(1w)
      • non-contrast CT Image

      All the numerical input data is extracted from the metadata fields of the patient.

      Output container files

      • report.pdf: PDF report.
      • NIfTI files (When available):
        • PREPROCESSED CT (preprocessed_ct.nii.gz): CT image preprocessed.
      • report/
        • online_report_baseline.html: Online report for baseline model, visible in the "Baseline model" tab.
        • online_report_24h.html: (When available) Online report for 24h model, visible in the "24h model" tab.
        • online_report_1w.html: (When available) Online report for 1 week model, visible in the "1w model" tab.
        • chart_baseline.png: Chart with the confidence for the baseline model.
        • chart_24h.png: (When available) Chart with the confidence for the 24 model.
        • chart_1w.png: (When available) Chart with the confidence for the 1w model.

      References

       

      Create free account now!

      Sign Up