Radiologist assisted machine learning
A computerized medical diagnostic system uses a training dataset that is updated based on reports generated by a radiologist. AI and/or CAD is used to make an initial determination of no finding, finding, or diagnosis based on the training dataset. Normal results with a high confidence of no finding are not reviewed by the radiologist. Low confidence results, findings, and diagnosis are reviewed by the radiologist. The radiologist generates a report that associates terminology and weighting with marked 3D image volumes. The report is used to update the training dataset.
1. A method comprising:
continuously updating a training dataset while analyzing medical image data with a medical image diagnostic computer having machine-learning capability, comprising the steps of:
using a medical image comprising a CT scan, a MRI scan, a PET scan, an ultrasound or a SPECT scan wherein said medical image comprises a first number of voxels;
using a segmentation algorithm to segment a structure within said medical image wherein said structure comprises a second number of voxels
wherein said second number of voxels is less than said first number of voxels, and
wherein said structure comprises multiple voxels corresponding to a sub-volume of said medical image,
wherein the selected sub-volume corresponds to an item on a diagnostic checklist;
analyzing the selected sub-volume to create a human-generated analysis; and
using the human-generated analysis of the sub-volume to update the training dataset.
2. The method of claim 1 comprising analyzing the selected sub-volume using the training dataset to create a machine-generated analysis with the diagnostic computer before manually analyzing the selected sub-volume.
3. The method of claim 2 comprising resolving disagreement between the human-generated analysis and the machine-generated analysis before using the human-generated analysis to update the training dataset.
4. The method of claim 3 comprising generating a computer-made explanation for the machine-generated analysis.
5. The method of claim 4 comprising updating the human-generated analysis based on the explanation before using the human-generated analysis to update the training dataset.
6. The method of claim 3 comprising prompting a consensus review of the human-generated analysis and machine-generated analysis.
7. The method of claim 6 comprising updating the human-generated analysis based on the consensus review before using the human-generated analysis to update the training dataset.
8. The method of claim 1 comprising the diagnostic computer retrieving and presenting patient-specific data pertinent to the item on the checklist to facilitate creation of the human-generated analysis.
9. The method of claim 1 wherein creating the human-generated analysis comprises creating at least one of: terminology describing findings or diagnosis; marked pixels or voxels; and an indication of certainty of the findings or diagnosis.
10. The method of claim 1 wherein creating the machine-generated analysis comprises creating at least one of: terminology describing findings or diagnosis; marked pixels or voxels; and an indication of certainty of the findings or diagnosis.
11. The method of claim 1 comprising filtering out tissue within the selected sub-volume that is not associated with a finding.
12. The method of claim 1 comprising using a three-dimensional cursor to select said segmented structure.
13. The method of claim 11 comprising automatically re-sizing the three-dimensional cursor to encompass tissue associated with the finding.
14. The method of claim 3 wherein the checklist comprises multiple items, each of which is analyzed, and comprising generating a report based on the human-generated analysis.
15. The method of claim 14 comprising including an indication of disagreement between the human-generated analysis and the machine-generated analysis.
16. The method of claim 1 comprising the three-dimensional cursor visually indicating confidence or dangerousness of a diagnosis.
17. The method of claim 1 comprising placing tissue associated with a finding in a virtual container.
18. The method of claim 17 comprising selecting a virtual container from a normal finding container, a disease-specific container, and differential diagnosis container.
19. An apparatus comprising:
a medical image diagnostic computer having machine-learning capability, the diagnostic computer comprising a non-transitory medium on which is stored computer program logic that continuously updates a training dataset while analyzing medical image data, comprising:
item selection logic that selects a sub-volume of a medical image corresponding to a structure by using a segmentation algorithm to segment said structure,
wherein the selected sub-volume corresponds to an item on a diagnostic checklist,
wherein said medical image comprises a CT scan, a MM scan, a PET scan, an ultrasound or a SPECT scan,
wherein said medical image comprises a first number of voxels,
wherein said structure comprises multiple voxels corresponding to said sub-volume of said medical image,
wherein said structure comprises a second number of voxels, and
wherein said second number of voxels is less than said first number of voxels;
input logic that receives input that creates a human-generated analysis of the selected sub-volume; and
update logic that updates the training dataset based on the human-generated analysis.
20. The apparatus of claim 19 comprising diagnostic logic that analyzes the selected sub-volume using the training dataset to create a machine-generated analysis before the human-generated analysis is generated.