(262 days)
INTContour provides a machine learning-based approach for the automatic segmentation of structures including treatment targets and organs at risk to support the radiation therapy treatment planning process. INTContour is intended as an initial method to segment and contour study series; therefore, this software must be used in conjunction with an appropriate software to edit the segmentation results if necessary. It is not intended to replace a thorough review by qualified medical professionals. INTContour is developed for use by dosimetrists, medical physicists, and radiation oncologists. The currently supported anatomical regions for automatic segmentation are head and neck, thorax, abdomen, and male pelvis.
INTContour is a software-only product that uses a machine learning-based approach to perform automatic segmentation of structures in medical images, coupled with tools for visualizing the segmentation results. A library of previously contoured expert cases serves as inputs to train the machine learning algorithms, specifically, convolutional networks (CNNs), to perform automatic segmentation. The results of the automatic segmentation will be stored in the DICOM Radiotherapy Structure Set (RTSTRUCT) format, which can be sent to desired destinations via the DICOM protocol. INTContour is intended to be used by dosimetrists, medical physicists, and radiation oncologists, and serves as an initial method to segment and contour study series. It must be used in conjunction with appropriate software to edit the segmentation results if necessary. The currently supported anatomical regions for automatic segmentation are head and neck, thorax, abdomen, and male pelvis. INTContour software is intended to be deployed within a hospital's private network on a workstation with an advanced graphics processing unit (GPU) and runs as a service. A web-based interface is used to access the service and manage the transfer of data, automatic segmentation, and visualization.
The acceptance criteria for Carina Medical LLC's INTContour device, along with its reported performance and details of the study proving it meets these criteria, are outlined below based on the provided document.
1. Acceptance Criteria and Reported Device Performance
The study used two primary metrics for evaluating the segmentation performance:
- Dice Similarity Coefficient (DSC): Used for larger organs.
- 95% Hausdorff Distance (HD95): Used for smaller organs.
The acceptance criteria were defined by comparing the performance of INTContour against a predicate/reference device (Smart Segmentation – Knowledge Based Contouring and AccuContour, respectively). The criteria were that INTContour's performance should be non-inferior to the predicate/reference device.
Table 1: Acceptance Criteria and Reported Device Performance
| Metric | Acceptance Criteria | Reported Device Performance |
|---|---|---|
| Dice Metric | The lower bound of the performance differences between INTContour and the predicate/reference device must meet or exceed the predefined threshold for all large organs. (Implies a minimum acceptable Dice similarity, demonstrating non-inferiority). | "By comparing the lower bound (Dice) or upper bound (HD95) of the performance differences between INTContour and the predicate/reference device with the threshold values, all organs have passed the acceptance criteria and demonstrated the noninferiority against the predicate/reference device." |
| 95% Hausdorff Distance (HD95) | The upper bound of the performance differences between INTContour and the predicate/reference device must meet or be below the predefined threshold for all small organs. (Implies a maximum acceptable Hausdorff distance, demonstrating non-inferiority). | "By comparing the lower bound (Dice) or upper bound (HD95) of the performance differences between INTContour and the predicate/reference device with the threshold values, all organs have passed the acceptance criteria and demonstrated the noninferiority against the predicate/reference device." |
Note: Specific numerical threshold values for Dice and HD95 were not provided in the document, only the statement that the INTContour met the non-inferiority criteria.
2. Sample Size and Data Provenance
- Test Set Sample Size: Not explicitly stated in terms of a numerical count of cases, however, the document notes: "Testing data was acquired from multiple sources than the training data that covers head and neck, thorax, abdomen, and male pelvis regions."
- Data Provenance:
- Country of Origin: Not specified.
- Retrospective or Prospective: Not specified, but the data was taken from "patients who went through radiation treatment," suggesting it was historical (retrospective) data.
- Patient Characteristics: Patients with ages 18-76, both male and female (implied by "various types of cancers"), and various types of cancers were included.
3. Experts for Ground Truth (Test Set)
- Number of Experts: "At least two trained personnel."
- Qualifications of Experts: Included "dosimetrist, medical physicist and/or radiation oncologist." Specific years of experience are not mentioned.
4. Adjudication Method (Test Set)
- The ground truth was performed by "at least two trained personnel... to minimize human bias in segmentation." This implies a consensus approach. However, the exact adjudication method (e.g., 2+1, 3+1, simple average, majority vote) is not explicitly detailed beyond "at least two trained personnel." It is not 'none' as there was a process involving multiple experts.
5. Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study
- No Multi-Reader Multi-Case (MRMC) comparative effectiveness study was performed or described. The study focused on the performance of the algorithm itself (standalone) and its non-inferiority compared to predicate devices, rather than measuring how human readers improve with AI assistance.
6. Standalone Performance Study
- Yes, a standalone (algorithm only without human-in-the-loop performance) study was conducted. The performance data section directly compares the "calculated metrics of INTContour against the predicate/reference device" using Dice and HD95, which are metrics for automated segmentation, not human-AI team performance.
7. Type of Ground Truth Used
- The ground truth for the test set was established through expert consensus based on manual segmentation by qualified medical professionals. Specifically, "Ground truth was performed by at least two trained personnel including dosimetrist, medical physicist and/or radiation oncologist."
8. Training Set Sample Size
- The exact sample size for the training set is not explicitly stated. The document mentions, "A library of previously contoured expert cases serves as inputs to train the machine learning algorithms, specifically, convolutional networks (CNNs)."
9. How Ground Truth for Training Set Was Established
- The ground truth for the training set was established from a "library of previously contoured expert cases." This implies manual contouring performed by experts, similar to the test set, though specific details about the number of experts or adjudication for the training set ground truth are not provided.
§ 892.5050 Medical charged-particle radiation therapy system.
(a)
Identification. A medical charged-particle radiation therapy system is a device that produces by acceleration high energy charged particles (e.g., electrons and protons) intended for use in radiation therapy. This generic type of device may include signal analysis and display equipment, patient and equipment supports, treatment planning computer programs, component parts, and accessories.(b)
Classification. Class II. When intended for use as a quality control system, the film dosimetry system (film scanning system) included as an accessory to the device described in paragraph (a) of this section, is exempt from the premarket notification procedures in subpart E of part 807 of this chapter subject to the limitations in § 892.9.