Sunteți pe pagina 1din 4

Integration Best Practices

Control Slider
Pearl’s Computer Vision Models provide a confidence score, between 0 and 100,
alongside each prediction. Confidence scores are a measure of the similarities between
the visual features & context of that detection and the corpus of data powering the AI; it
is not a measure of certainty in the manner a person would say “I’m 60% confident”.
These confidence scores are useful to provide different modes of operation to the end
user.

Our recommendation is to establish 3 positions for toggling results:

• Default Display predictions above a mid-range confidence score.


This serves to strike a balance between capturing areas
of concern without overwhelming the dentist with too
much data at once.

• Sensitivity Boost Display predictions above a lower confidence score


cutoff. Additional predictions shown here serve to
provide more options for consideration by the dentist.

• Specificity Boost Raise the confidence score cutoff to display fewer


predictions. Decreased predictions serve to provide only
the most pronounced cases to the dentist.

This can be controlled by the end user via a slider:

Fig. 1 Example of Slider


Rendering Context
To increase readability of the results it is recommended to alter the styling of the
predictions depending on score; highlighting higher scores & downplaying visibility of
lower. The styles should relate to the confidence score cutoffs for the toggled positions
with prediction being stylized according to the highest score tranche to which they
belong (Default should present 2 styles; Specificity boost 1 style, Sensitivity boost 3
styles).

Potential aspects to alter:


• Color of Box
• Stroke Thickness
• Stroke Style

Fig. 2 Examples of Modifications

The label denoting the category of prediction should be clearly associated to its box.
This is typically achieved via a hover state label.

User-Controlled Hiding of Results

Given the wide range of detections Pearl’s Computer Vision models provide it is highly
recommended users are given the ability to filter which types of results are presented.

To achieve this, we recommend providing a mechanism to Hide & Show detections:

• All Detections
• Pathologies
• Non-Pathologic Detections

Fig. 3 Typical Show-Hide Iconography


An alternative path is to provide a mechanism to click through different view ports each
representing a type of classification, such as:

• Caries
• Margin Discrepancies
• Existing Crowns
• Existing Fillings
• Additional Existing Restorations

Fig. 4 Example Viewport Selector

Recommended Workflow
Our recommended workflow for interacting with the API is as follows:

1. POST to /login to receive a token for all subsequent requests (TTL: 24 hrs)
a. Token must be sent in a Bearer authorization style header i.e.
Authorization: “Bearer {token}”
2. POST to /images/uploadUrl
a. Gets a pre-signed request to upload your image
3. POST to the url supplied in 2a with the image itself to upload it for processing
4. POST to /images
a. Kicks off the processing once the image has been uploaded
5. GET /images
a. Poll for your responses from our detectors. In general processing takes
less than 10 seconds per image, we recommend polling with a GET
request once per second after the POST request to retrieve annotations
for the image. When the “state” object of the image is set to
“isCompleted”: true, polling can cease as no further work is being done
processing the image.

Bounding boxes can be rendered based on the coordinates in the contour element. X &
Y are pixel values for the upper left corner of the box with height and width
representing the pixel size of the bounding box.
Recommended Confidence Scores
These scores have been generated based on clinician performance review. Please note
that while Pearl AI is intended for use with Periapicals & Bitewings not all categories are
supported for both types of radiograph.

Category Sensitivity Boost Default Specificity Boost

Caries
.2 .7 .8
(Bitewing Only)
Margin
.2 .7 .9
Discrepancy
Crown .3 .5 .9
Implant .1 .2 .6
Root Canal .1 .2 .6
Bridge .0 .0 .0
Filling .2 .4 .7

UX Examples:
GIF of Coloration of boxes & control slider:
https://drive.google.com/file/d/1xFBZCkGzfsdEgdrxGe-pRvfVlg-VxMrT/view

Video of coloration, slider, & hover labels: https://drive.google.com/file/d/1--


HQFVsnwXryrRC_KsqZTMbuckdpup5U/view

S-ar putea să vă placă și