Search Results

IRISeg is intended for use as a software application that receives DICOM compliant MR or contrast-enhanced CT images, provides manual and machine learning-enabled tools for image analysis and segmentation, and creates an output file that can be used to render a 3D model for preoperative surgical planning and intraoperative display. The use of IRISeg may include the generation of preliminary segmentations using machine learning algorithms. IRISeg is intended for use by qualified professionals. The output file is meant for visual, non-diagnostic use and shall be reviewed by clinicians who are responsible for all final patient management decisions.

The machine learning enabled kidney CT auto-segmentation tool is intended for use for adult patients with contrast-enhanced, axial kidney CT images with slice thickness 3mm or less.

IRISeg is a standalone software application created by Intuitive Surgical for segmentation of CT and MR images and generation of output files that can be rendered as virtual 3D models of anatomical structures. IRISeg is designed to provide qualified professionals ("users") with a machine learning (ML)-based tool for auto-segmentation of kidney anatomy based on CT scans and non-ML manual tools for segmentation based on CT and MR scans.

Note that there have been no changes to existing tools or introductions of new tools between the predicate and subject devices.

Input File
IRISeg can open and load CT or MR imaging files in DICOM (Digital Imaging and Communications in Medicine) format, and segmentation label files in NIfTI (Neuroimaging Informatics Technology Initiative) format from an accessible storage location.

Output File
Following the use of IRISeg to segment CT or MR imaging files, the software can be used to generate an output file that can be used to render virtual segmented 3D models.

IRISeg Manual Tools
IRISeg includes a variety of tools for users to manually edit segmentation labels, such as Paintbrush tools, Eraser tools, Connected Component Selection, Free Curve Selection, Morphological operations, Mathematical Operations.

Manual tools alone can be used to manually segment (annotate) CT and MR scans.

Manual tools can also be used to modify the output of the ML-based auto-segmentation algorithm. The ML-based auto-segmentation does not generate mass labels. Users must segment and label renal masses using manual tools.

IRISeg ML-Based Auto-Segmentation Tool
IRISeg includes an ML-based auto-segmentation algorithm (cleared under K242461 and unchanged in the subject device) for automatic segmentation of four kidney structures from CT imaging. The auto-segmentation algorithm is a neural network based ML algorithm. It is trained on segmented kidney CT models that were sourced from clinical data processed during commercial operation of the cleared IRIS 1.0 system (K182643). Each 3D model was reviewed by one U.S board certified radiologist. The input is a CT image (series of 2D slices). The output of the model is four probability maps for kidney parenchyma, kidney artery, kidney vein, and collecting system. The probability maps are thresholded to generate binary masks for kidney parenchyma, kidney artery, kidney vein and collecting system. The ML-based auto-segmentation does not generate binary masks for kidney masses.

The algorithm output is intended as an initial estimate of the segmentation. The user must use the manual tools to update the initial algorithm output to generate the kidney CT 3D model.

The development of the IRISeg kidney CT ML-based auto-segmentation algorithm followed FDA's Good Machine Learning Practices for Medical Device Development: Guiding Principles, October 2021.

This 510(k) clearance letter pertains to IRISeg, a software application that assists in the segmentation of CT and MR images to create 3D models for surgical planning. The document heavily references a predicate device (K242461) as the software itself (including the ML-based auto-segmentation algorithm) has not changed. The clearance addresses the expansion of the indications for use to include MR images for manual segmentation and the standalone nature of the software.

Here's a breakdown of the requested information based on the provided text:

1. Table of Acceptance Criteria and Reported Device Performance

The provided document doesn't explicitly state quantitative acceptance criteria for the subject device (K251763) beyond "all tests met the acceptance criteria" in reference to general software testing. However, it does state that the ML auto-segmentation algorithm was "not modified in the subject device, and therefore the performance of the ML algorithm is as effective as in the predicate device."

To address the request, we would need to infer information from the predicate device's clearance (K242461), which is not fully detailed in this document. Since the ML algorithm remains the same, any performance metrics from K242461 would be applicable. However, without details from K242461, we can only report what is explicitly mentioned here about K251763.

The document notes that the "ML-based auto-segmentation does not generate mass labels" and "Users must segment and label renal masses using manual tools." This is a limitation, not a performance metric, but relevant to the overall utility.

2. Sample Size Used for the Test Set and Data Provenance

For ML Auto-Segmentation (performance inherited from K242461):

• Test Set Sample Size: Not explicitly stated for K251763 or K242461. The document only mentions the training data for the ML model.
• Data Provenance: Not explicitly stated for the test set.

For Manual Segmentation and General Software Testing (K251763):

• Test Set Sample Size: Not explicitly stated.
• Data Provenance: Not explicitly stated.

The document indicates that the machine learning model was trained on "clinical data processed during commercial operation of the cleared IRIS 1.0 system (K182643)." This suggests real-world clinical data, likely retrospective. The country of origin is not specified but given "U.S. board certified radiologist," it's reasonable to infer the data includes U.S. clinical data.

3. Number of Experts Used to Establish Ground Truth for the Test Set and Qualifications

For ML Auto-Segmentation (training data from IRIS 1.0, K182643):

• Number of Experts: "one U.S board certified radiologist" per 3D model.
• Qualifications: U.S. board certified radiologist. Years of experience are not specified.

For the test set(s) used for K251763 (functional, usability, cybersecurity, and manual segmentation for MR):

• Number of Experts/Users for Ground Truth: Not explicitly stated. The document refers to "qualified professionals" as intended users, who would perform manual segmentation tasks.

4. Adjudication Method for the Test Set

For ML Auto-Segmentation (training data ground truth from IRIS 1.0, K182643):

• Adjudication Method: "Each 3D model was reviewed by one U.S board certified radiologist." This implies a single-reader ground truth without explicit adjudication unless that review process involved an internal review cycle. No 2+1 or 3+1 method is mentioned.

For the test set(s) used for K251763:

• Adjudication Method: Not explicitly stated for any of the general software testing.

5. Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study

• Was an MRMC study done? No, an MRMC comparative effectiveness study is not mentioned or described in the provided text.
• Effect size of improvement: Not applicable, as no MRMC study was conducted or reported. The document focuses on the standalone algorithm's performance and the general software functionality.

6. Standalone (Algorithm Only Without Human-in-the-Loop) Performance

• Was standalone performance done? Yes, implicitly. The document states: "The output of the model is four probability maps for kidney parenchyma, kidney artery, kidney vein, and collecting system. The probability maps are thresholded to generate binary masks..." This describes the direct output of the ML algorithm.
  • It also clarifies: "The algorithm output is intended as an initial estimate of the segmentation. The user must use the manual tools to update the initial algorithm output to generate the kidney CT 3D model." This indicates that while standalone output exists, it is expected to be refined by a human.
  • Performance for "Machine Learning Auto-Segmentation Testing" is specifically mentioned under K242461, and stated to be the "Same" for K251763. However, specific performance metrics (e.g., Dice coefficient, sensitivity, specificity) for this standalone performance are not provided in this document.

7. Type of Ground Truth Used

For ML Auto-Segmentation Training Set:

• Type of Ground Truth: Expert consensus (specifically, review by "one U.S board certified radiologist" per model). This is a form of expert annotation/segmentation.

For the test set(s) of K251763:

• Type of Ground Truth: Not explicitly stated for the general software testing. For manual segmentation, it would likely involve expert-derived reference segmentations or user-defined targets.

8. Sample Size for the Training Set

• Sample Size for Training Set: Not explicitly stated. The document mentions training on "segmented kidney CT models that were sourced from clinical data processed during commercial operation of the cleared IRIS 1.0 system (K182643)," but the number of such models is not provided.

9. How the Ground Truth for the Training Set Was Established

• How Ground Truth Was Established: "Each 3D model [used for training the ML-algorithm] was reviewed by one U.S board certified radiologist." This indicates that a U.S. board-certified radiologist manually segmented or reviewed and confirmed the segmentation of the 3D models used as ground truth for training.

Ask a specific question about this device

IRISeg is intended for use as a software application that receives DICOM compliant contrastenhanced CT images, provides manual and machine learning-enabled tools for image analysis and segmentation, and creates an output file that can be used to render a 3D model for preoperative surgical planning and intraoperative display. The use of IRISeq may include the generation of preliminary segmentations using machine learning algorithms. IRISeg is intended for use by qualified professionals. The output file is meant for visual, non-diagnostic use and shall be reviewed by clinicians who are responsible for all final patient management decisions.

The machine learning enabled kidney CT auto-seamentation tool is intended for use for adult patients with contrast-enhanced, axial kidney CT images with slice thickness 3mm or less.

IRISeg Version 3.1 ("IRISeg") is a standalone software application created by Intuitive Surgical for segmentation of contrast-enhanced kidney CT images and generation of output files that can be rendered as virtual 3D models of kidney organs. IRISeg is designed to professionals ("users") with manual and machine learning (ML)-based tools for segmentation of kidney anatomy based on CT scans.

Let's break down the acceptance criteria and study details for the IRISeg device, based on the provided FDA 510(k) summary.

Device Name: IRISeg
Intended Use: Software application for image analysis and segmentation of contrast-enhanced kidney CT images, creating an output file for rendering a 3D model for pre-operative surgical planning and intraoperative display. The machine learning enabled kidney CT auto-segmentation tool is intended for adult patients with contrast-enhanced, axial kidney CT images with slice thickness 3mm or less. The output is for visual, non-diagnostic use and requires review by clinicians.

1. Table of Acceptance Criteria and Reported Device Performance

The document characterizes performance using the Sørensen–Dice coefficient (DSC) for segmentation accuracy of anatomical structures and Mean Distance to Agreement (MDA) for the collecting system. While explicit "acceptance criteria" are not given as numerical cutoffs, the reported performance demonstrates the algorithm's accuracy relative to expert consensus.

(Note: The document states "test results showed that all tests met the acceptance criteria" generally for software V&V, and then presents these quantified performance metrics for the ML algorithm. We infer that these reported intervals fulfill the internal, unstated numerical acceptance thresholds for clinical performance.)

2. Sample Size Used for the Test Set and Data Provenance

• Test Set Sample Size: 81 kidney CT scans.
• Data Provenance: The document states that the clinical data for training was "sourced from clinical data processed during commercial operation of the cleared IRIS 1.0 system (K182643)." For the test set, it explicitly states "No imaging study used to verify performance was used for training; independence of training and testing data were enforced at the level of the scanning institution, namely, studies sourced from a specific institution were used for either training or testing but could not be used in testing ensured diversity in patient population, scan parameters and scanner manufacturers." This implies a retrospective study using existing clinical CT scans. The country of origin is not explicitly stated, but given the FDA submission and the use of "U.S board certified radiologist," it can be inferred that the data is likely from the U.S. or at least relevant to U.S. clinical practice. The data is retrospective.

3. Number of Experts Used to Establish the Ground Truth for the Test Set and Qualifications

• Number of Experts: Three (3) U.S. Board Certified Radiologists.
• Qualifications: U.S. Board Certified Radiologists. (No further details on years of experience or subspecialty provided beyond "Board Certified").

4. Adjudication Method for the Test Set

The ground truth was established by the "consensus of three U.S Board Certified Radiologists." This implies a form of consensus-based adjudication, but the specific mechanics (e.g., majority vote, sequential review with final agreement, arbitration by a fourth reader) are not detailed beyond "consensus."

5. Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study

No, an MRMC comparative effectiveness study involving human readers with vs. without AI assistance was not reported in this summary. The AI performance was evaluated in a standalone manner against expert consensus ground truth. The device output is explicitly non-diagnostic and requires review by clinicians, emphasizing a human-in-the-loop workflow, but the study focused on the algorithm's raw segmentation accuracy.

6. Standalone (Algorithm Only Without Human-in-the-Loop Performance) Study

Yes, a standalone performance study was conducted. The performance metrics (DSC and MDA) characterize the ML algorithm's output directly when compared to expert-derived ground truth. The summary states, "Performance was evaluated by comparing segmentations generated by the kidney CT machine learning algorithm against segmentations generated by a consensus of three U.S Board Certified Radiologists for the same imaging study." This is a direct measurement of the algorithm's standalone accuracy.

7. Type of Ground Truth Used

The ground truth used was expert consensus by three U.S. Board Certified Radiologists, based on their segmentation of the medical images.

8. The Sample Size for the Training Set

The sample size for the training set is not explicitly stated in the provided summary. It states the model was "trained on segmented kidney CT models that were sourced from clinical data processed during commercial operation of the cleared IRIS 1.0 system (K182643)."

9. How the Ground Truth for the Training Set Was Established

The ground truth for the training set was established by review from one U.S. board certified radiologist per 3D model. The summary states: "Each 3D model was reviewed by one U.S board certified radiologist."

Ask a specific question about this device