Machine Learning (ML) is now able to help medical professionals diagnose important cases such as diabetes mellitus and metastatic breast cancer.
No matter how much the algorithm has improved, there useless unless the doctor is able to understand the data.
So how do doctors communicate with information, can be important determinants in the usefulness of ML technology?
It can be an image of a cancer scan or blunt trauma. Modern-day technology enables skimming the computer through thousands of images and retrieves the closest information to the query.
For example, reverse image search options in Google Images
Now, ‘Digital Pathology’ is being widely adopted. Doctors can now check the data obtained on the computer as images.
And, along with it, you can gain insight while making the entire process of diagnosis easier.
On these lines, a group of Google Artificial Intelligence researchers has joined the digital pathology to present machine learning tools and methods:
Similar Image Search for Histopathology SMILY
Human-oriented tools to combat incomplete algorithms while making medical decisions
The first step to developing SMILY was to implement an deep learning model, which was trained using 5 billion natural,
non-pathology images (for example, dogs, trees, man made objects etc.).
During training, learn to distinguish between networks by comparing embedding with similar images and different images.
Embedding is a summary of numerical vectors, which is a compression of all the pictures played in the deep learning model during training.
SMILY allows a user to select an area of interest and visually obtain similar matches.
As can be seen in the example below, where by selecting a small area in a slide,
SMILY will discover the database of billions of trillion images efficiently in a few seconds.
Because pathology images can be seen at different magnification (zoom level), SMILY automatically searches images as input image on the same magnification.
Although these machine learning tools tend to greatly reduce the time spent by traditional methods for traditional image retrieval,
they fall short in understanding the user’s or intent to search.
A doctor can see the same report for different things and it varies in different cases.
In order to make SMILY more interactive, researchers have refined the device to enable end users, which explains what similarity means:
Refine-by-area image, the pathologist allows the crop in the area of interest, which limits the search of that area.
Refine-by-example gives users the ability to choose a subset of search results and get more results like those
Refinery-concept slider can be used to tell whether there is a more or less clinical concept in search results (for example, fused, gland)
The above concepts should not be made in the machine learning model. Instead, this method allows end users to create new concepts,