Search Results

Hermes is intended for use as a software interface and image segmentation system for the transfer of medical imaging information to an output file. It is also intended as preoperative software for diagnostic and surgical planning.

For these purposes, output files can also be used for the fabrication of physical replicas using traditional or additive manufacturing methods. The physical replicas can be used for diagnostic purposes in the field of orthopedic applications.

The software is intended to be used in conjunction with other diagnostic tools and expert clinical judgment.

Vent Hermes takes DICOM based CT scans from the scanning machine and plots the data generated by the machine in the form of slices of data, similar to all commercially available predicates.

The intensity data (Hounsfield Unit data) inside these is displayed to ensure that all intensity data is visible to the users. The point clouds from these data points are used to find landmarks in discrete locations, unlike the predicates that use bone outlines from automatic or semi automatic or manual sources to generate points for landmarks. This approach allows for more accurate predictions, as validated in previous studies by radiology experts. Published literature is used to define the 3 cut planes where proximal tibia and distal and posterior femoral cuts are calculated.

The product does not currently make any claims of individual implant fits but planes shared by all surgeon and implant manufacturers as the first step. The python code has been history file with design terations recorded on all major changes to the system. The segmentation aspect is a Convolutional Networks model trained with arthritic CT scans from 120 de-identified patients with varying arthritic states, demographics, and centers partnering with surgeon champions.

The seqmentation and landmarking system was validated aqainst manual, Materialize Mimics 24 autosegmentaiton, and Synopsys Simpleware knee auto segmentation models, and results were reviewed with the Chief of Radiology of Hospital for Special Surgery in New York, NY. The results were published in the International Society for Technology in Arthroplasty 2022 annual conference. Vent Hermes is a cloud-based python code that allows users to upload the de-identified DICOM files to their encrypted Google Cloud Platform folder (a business agreement certified partner with HIPAA and 21 CFR 11 certifications available).

Once the files are uploaded, the cloud server de-identifies the folder again for redundant safety, then passes it to Vent Creativity's internal server for the system to automatically run the scripts to generate the bone models, landmarks, and the final cut planes and returns these files to the user file same file number. Once the point clouds, meshes, and landmarks are calculated, any number of CAD or visualization tools can be used to review the data.

The system uses a custom Ul qenerated on the OHF medical imaging platform with system version controls and validation documents on file. This portion is considered to be software outside of the medical device as no calculation is allowed outside the python code. The visualizations are final and are for user review and approval only.

Here's a breakdown of the acceptance criteria and the study that proves the device meets them, based on the provided text:

Acceptance Criteria and Reported Device Performance

Study Details

1. A table of acceptance criteria and the reported device performance: (See above table)

2. Sample size used for the test set and the data provenance:

   • Test Set Sample Size: Implicitly, the validation was done on the same data used for training the segmentation model, which was "arthritic CT scans from 120 de-identified patients." However, specifically for the validation against predicates, the text refers to a "study compared the segmentation accuracy of Hermes with two predicate segmentation tools, Synopsys and Mimics, using the Dice Similarity Coefficient (Dice score) as the primary metric." It doesn't explicitly state if this comparison was on all 120 patients or a subset.
   • Data Provenance:
     • Country of Origin: Not explicitly stated, but the mention of "Hospital for Special Surgery in New York, NY" suggests data from the US.
     • Retrospective or Prospective: Retrospective, as it refers to "de-identified patients" and "arthritic CT scans from 120 de-identified patients."

3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts:

   • Number of Experts: Not explicitly stated for establishing ground truth for the specific test set.
   • Qualifications of Experts: The "results were reviewed with the Chief of Radiology of Hospital for Special Surgery in New York, NY." This implies a highly qualified expert. For the training data, "surgeon champions" were involved, but their role in establishing ground truth for training vs. clinical input is not detailed.

4. Adjudication method (e.g., 2+1, 3+1, none) for the test set:

   • Adjudication Method: Not explicitly stated. The comparison was primarily quantitative (Dice score) against predicate auto-segmentation models (Materialize Mimics 24 autosegmentation, Synopsys Simpleware knee auto segmentation models) and "manual" segmentation. The involvement of the Chief of Radiology appears to be for review and confirmation of results, rather than a formal adjudication process amongst multiple readers for ground truth generation.

5. If a multi-reader multi-case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance:

   • MRMC Study: No, an MRMC study comparing human readers with and without AI assistance was not done. The study compares the algorithm's performance to other algorithms and "manual" segmentation benchmarks, primarily focusing on segmentation accuracy.

6. If a standalone (i.e., algorithm only without human-in-the-loop performance) was done:

   • Standalone Performance: Yes, standalone performance was evaluated through the "Segmentation Analysis and DICE Score Study," comparing Hermes's segmentation with Synopsys and Mimics predicate tools. The system is described as automatically running scripts to generate bone models, landmarks, and cut planes.

7. The type of ground truth used (expert consensus, pathology, outcomes data, etc.):

   • Type of Ground Truth: The text mentions "validated against manual" segmentation, which implies manual segmentation (likely performed by an expert or experts) served as a form of ground truth for comparison. This is often an expert-driven ground truth. It also compares against the output of other established auto-segmentation software (Mimics and Synopsys).

8. The sample size for the training set:

   • Training Set Sample Size: "120 de-identified patients with varying arthritic states, demographics, and centers."

9. How the ground truth for the training set was established:

   • Ground Truth Establishment for Training: The text states the segmentation aspect is a "Convolutional Networks model trained with arthritic CT scans from 120 de-identified patients." It does not explicitly state how the ground truth labels for these 120 patients were established for the training process (e.g., if they were manually segmented by experts, or derived from other clinical data). It only mentions that "surgeon champions" were involved in "partnering with" centers from which the data came, implying clinical input, but not specifically delineating their role in creating the ground truth annotations for training. Given the "validation against manual" for the test set, it's highly probable the training data also utilized manual segmentations as ground truth.

Ask a specific question about this device