A new self-teaching artificial intelligence (AI) tool has been developed to aid in the diagnosis and treatment prediction in rare diseases.
By virtue of scant information on them, rare diseases present a challenge for clinicians both in their identification and in determining the best course of treatment.
Artificial intelligence, or machine-learning tools, are emerging as useful in assisting the experience and critical thinking of human clinicians. AI tools have been shown to be effective in diagnosis, prognosis and treatment predictions in epilepsy studies, looking for potentially cancerous lung nodules and the care of patients with traumatic brain injuries.
There are understandable concerns, however, about the implementation of AI in medicine because AI is only as good as the data that it is fed – so it may reflect certain biases in demographics. It’s also very new and not well legislated.
But AI experts and clinicians alike are confident that machine-learning algorithms will play a role in patient care in the near future, not as replacements for human doctors, but by complementing human experience and wisdom to maximise our ability to help patients.
Read more: Artificial intelligence tool swings into action, discovering alternative physics variables!
Researchers from the machine learning-focused Mahmood Lab at Brigham and Women’s Hospital in Boston, Massachusetts, have developed a deep learning algorithm which can teach itself how to identify similar features in large pathology image repositories to help diagnose and generate treatment guides for rare diseases. Their results are published in the journal Nature Biomedical Engineering.
Known as SISH (Self-Supervised Image search for Histology), the new tool is a type of “self-teaching” algorithm. At its most basic, deep learning attempts to imitate the complex neural networks in our own brains through algorithms. These algorithms can then “learn” things about data sets by finding patterns and trends, much like we do in our daily experience.
SISH acts like a search engine for pathology images. Among its many potential applications, it has proven adept at identifying rare diseases and helping clinicians determine which patients are likely to respond to certain therapies.
“We show that our system can assist with the diagnosis of rare diseases and find cases with similar morphologic patterns without the need for manual annotations and large datasets for supervised training,” says senior author Dr Faisal Mahmood, from the Brigham’s Department of Pathology. “This system has the potential to improve pathology training, disease subtyping, tumour identification, and rare morphology identification.”
Given AI has been around for a little bit now, it’s not surprising that other tools have been tested for these kinds of uses.
Fundamentally, they all rely on the data – and modern electronic databases can store an immense number of digital records. A significant proportion of this data comes in the form of images. In pathology, these are particularly whole slide images (WSIs) which are the complete scans of a microscope slide, creating a single high-resolution digital file.
However, these high-fidelity images can be large files. As more of them fill digital repositories, searching through WSI databases can be time consuming, and computationally complex and expensive.
Brigham researchers overcame this issue with SISH which teaches itself to recognise features and find analogous cases in databases at a constant speed regardless of the size of the database.
The pathologists and AI experts tested the speed and ability of SISH to correctly retrieve information for both common and rare cancers.
SISH was able to successfully obtain images accurately and at high speed from a database of tens of thousands of WSIs from 22,000 patient cases. Over 50 different disease type and more than a dozen anatomical sites were represented in the data set.
The new algorithm outperformed other methods in many scenarios, including in identifying disease subtype. Of particular importance, was SISH’s ability to maintain a constant search speed even as the databases expanded in size, and when using diverse data sets.
Read more: Promise and problems: how do we implement artificial intelligence in clinical settings and ensure patient safety?
A drawback, however, is that SISH does require a huge amount of memory. The new tool also has limited context awareness when it comes to WSIs in large tissues, and it is currently only useful in identifying single images.
But the researchers are confident their new tool represents a development in the proficiency of rare disease diagnosis and analysis.
“As the sizes of image databases continue to grow, we hope that SISH will be useful in making identification of diseases easier,” said Mahmood. “We believe one important future direction in this area is multimodal case retrieval which involves jointly using pathology, radiology, genomic and electronic medical record data to find similar patient cases.”