How should I interpret the results?
Interpreting the results requires careful consideration of the image quality and the context in which it was taken. By following the recommended guidelines, you can ensure more accurate and reliable results.
The following questions frequently arise when users interpret the results.
At what threshold should I consider the probability of a condition to be high enough to make a diagnosis?
This way of thinking is not the best way to conceptualise the issue.
The conclusions object of the device returns an array of possible conditions. But you must keep in mind that conclusions is the output of a classification algorithm. In other words, the algorithm knows a list of around 300 classes, and when it looks at an image, it distributes across the 300 classes.
As a result, a condition may have a probability of , and another one . The important thing is that, if you sum the probability of all conditions, you will get , because it is a distribution.
With this in mind, it may be true that, if a condition has a probability higher than , in a large number of cases, that will be the actual condition. But that is not the correct methodology of analysing a distribution.
Classifying and diagnosing are different things. To diagnose, you must classify; but it is a separate process. Diagnosing means looking at the distribution of the probability, but also considering the context and understanding the relationship between the conditions.
In this regard, the device contains keys that utilise the array of conclusions and have clinical knowledge embedded. For instance, isMalignantSuspicion combines the probabilities of conditions that are considered malignant.
You can find more information about this on the Workflows section of the documentation.
How important is the image?
The image is the most important factor to consider when interpreting the results, because the output of the device is absolutely and directly influenced by the image.
High-quality images ensure more accurate and reliable results. The ability to detect and assess conditions is significantly enhanced by clear, well-focused, and properly lit images. It is crucial to remember that these outcomes are directly influenced by the quality of the analyzed image.
We strongly recommend reviewing our How to take pictures guide for essential tips on optimal image capture, including using natural lighting and maintaining a steady hand.
For a comprehensive understanding of the results, their interpretation, and the variability depending on the context, we recommend consulting the Workflows section of the documentation.
What does non-specific lesion mean?
A non-specific lesion result indicates that the device has not identified distinctive features that match any specific condition. This outcome occurs when the system cannot confidently classify the lesion based on the image provided.
This usually happens for two reasons:
- The image does not contain a condition.
- The condition contained in the image is not one of the conditions that the device can identify.
- The lesion is so small or insignificant in the image that the device does not see the lesion or does not know what to look at. Several factors can contribute to this, such as the quality of the image, the positioning of the camera, or the inherent characteristics of the lesion itself.
The 2nd and 3rd options are actually the same as the 1st. For the device, a condition that is not visible is the same as no condition, because it only knows what the pixels of the image are showing.
It is important to note that a "non-specific lesion" result does not necessarily imply the absence of a lesion; rather, it indicates that, based on the provided image, the AI could not make a definitive assessment.
How common is non-specific lesion as an output?
Understanding how frequently the class non-specific lesion appears in the output can be challenging, because it depends on the user's input more than anything else.
The non-specific lesion result usually appears when there is not a dermatological condition present. For instance, if an organisation offers patients to upload images of anything they want, they may end up uploading selfies of the face, with cosmetic "issues". In this case, the device will correctly say that the face is not a lesion by outputting a non-specific lesion result.
Otherwise, the non-specific lesion result often appears when the lesion is either absent or occupies a small portion of the image. In such cases, the device will not be able to identify the lesion accurately, leading to a non-specific lesion result.
What do low-probability conditions mean?
When the top conditions in a results shows a low probability, it indicates that the probability of the detected condition for the given image is low. In other words: the device has not clearly identified distinctive features that match any specific condition.
A low probability score, such as 10% or 15%, means that the software is not sure about it's own output. Consider repeating the test with a better image for more accurate results.
It is important to note that a result with low probability does not necessarily imply the absence of a lesion; rather, it indicates that, based on the provided image, the device could not make a definitive assessment.
How can I increase the probabilities in the results?
You can follow these steps to increase the probabilities in the results:
1. Retake the picture following the instructions
To ensure accurate analysis and classification of lesions by the device, it is essential to capture high-quality images where the lesion is the focal point and occupies a significant portion of the frame. Blurry or poorly focused images can lead to inconclusive results, and if the lesion is too small or distant, it difficult for the software to analyze accurately.
We recommend reviewing our How to take pictures guide for essential tips on optimal image capture, including using natural lighting and maintaining a steady hand. If you receive a result with Non-Specific Lesion or a low probability in top conditionss, consider retaking the picture with a higher-resolution camera to capture more detail, thereby improving the device's ability to make a confident diagnosis.
Alternatively, you can simply crop the image to focus on the area containing the lesion, removing any surrounding elements or artifacts that might interfere with the analysis. This adjustment can help enhance the accuracy of the device's assessment by ensuring that the lesion is the primary focus of the image.
2. Improve instructions for users
Depending on the integration chosen, you can refine the guidance provided to individuals capturing the images. For instance, with the Iframe integration you have the option to configure the iframe to display instructions before a photo is uploaded. This ensures that users are equipped with the necessary information to capture the highest quality images possible.
3. Improve the workflow
Reach out to the Legit.Health team and we will be happy to work with you to improve the workflow to your specific objectives. This collaboration can include utilizing dashboards that provide analytics on conditions, enabling you to make informed decisions and adjust processes based on your goals.