Commercializing AI in Medical Imaging

AI has been used in medical imaging for a while now. Dozens of well funded companies operate in this space. Although the new machine learning techniques, massive compute power and rise is electronic medical images has created many impressive solutions, commercialization is a whole different ballgame. In this article I hope to review how multiple companies are commercially positioning themselves in this space. Mainly what are the common commercialization pathways and if it can lead us to sort of a playbook.

Since last one decade the key focus of AI application in medical imaging has remained mostly on the diagnostic. Either computer-aided diagnostic or more independently executing algorithm doing at least as good of job as human reviewers. We can call this as the first generation of AI on medical imaging products and companies. In last few years however we see companies going beyond disease diagnostic and thinking hard about better integration with end-to-end workflow, creating what we can call as the next generation of companies.

Let's start with PathAI, a company started in 2016 who has built a platform to enable substantial improvements to the accuracy of diagnosis. The company has expanded into the measurement of therapeutic efficacy for complex diseases, leveraging the recent advancements in machine learning. The company raised $165 million is 2021 and also acquired Poplar Healthcare, a laboratory services company.

They started with a narrow scope of how can you identify cancerous cells in histopathology images. [?] And now have some impressive algorithms which can help you identify other relevant biomarkers like PD-1 expression on tumor or immune cells at least on par with human pathologists. One more cool thing about Path.ai is they hav developed a platform to collect exhaustive annotations (of PD-L1 positivity in this case) from a crowdsourced network of pathologists for analytic validation.

In September, in a historic milestone, the FDA approved the first ever AI-based pathology product by Paige.ai for clinical use. Paige Prostate is the first and only AI-based pathology product to receive FDA approval for in vitro diagnostic use in detecting cancer in prostate biopsies. Paige Prostate is a clinical-grade AI solution designed to identify foci that could indicate cancer, enabling fast, accurate in vitro diagnosis. In the clinical study submitted to the FDA, pathologists using Paige Prostate were shown to increase over 7 percentage points in sensitivity in correctly diagnosing cancer (from 89.5% to 96.8%). Pathologists using Paige Prostate had a 70% reduction in false negative diagnoses and a 24% reduction in false positive diagnoses.

Although PaigeAI and PathAI are destined to compete with each other, they seem focussed on generating their own value differentiators. Paige is primarily after diagnostics through FDA-approval path. Although Path has companion diagnostics as part of their core value proposition they seem to be focussed on offering variety of solutions like 

• Identify new insights on pharmacodynamics, mechanism of action, and patient stratification
• Quantify biomarkers for patient enrollment (quantify HER2 expression, and measure stain intensity, artifact content, tumor area, and DCIS)
• Surrogate endpoint prediction (e.g. MPR in NSCLC)
• Correlative studies in prospective clinical trials
• Measuring efficacy of treatments through various strategic partnerships. 

In December 2021, Gleamer AI demonstrated that their algorithm can quickly detect and flag X-rays with positive fractures helping hospitals reduce missed fractures by 29%. Due to the large number of X-rays that have to be read by radiologists, patients often have to wait hours in the ER before they can be seen, evaluated, and receive treatment. Fracture interpretation errors represent 24% of harmful diagnostic errors in the ER.

Dental industry is also seeing increased interest in companion diagnostics. Overjet a startup spun off from Harvard's innovation lab went through series A and series B in 2021 and now values at 425 million. Started in 2018 Overjet has been hyperfocussed on execution. In May of last year, FDA has authorized one of the first artificial-intelligence-based technologies for use in dentistry, a software platform from Overjet. With the 510(k) clearance, Overjet will be able to market its Dental Assist software for clinical use, selling it directly to dental practices. The software is already employed by insurance companies to make claims processing more accurate and efficient. 

Generation 2

The successful use of artificial intelligence (AI) for diagnostic purposes has prompted the application of AI-based cancer imaging analysis to address other, more complex, clinical needs. For example using the radiology images can solve,  

• prognostication of outcome across multiple cancers
• prediction of response to various treatment modalities
• Discrimination of benign treatment confounders from true progression
• Identification of unusual response patterns and prediction of the mutational and molecular profile of tumours.

AI powered workflow

Finally, my favorite example of Viz.ai. This company is a textbook case of how AI needs to be seamlessly incorporated into medical workflow. In 2020, Viz.ai came up with a workflow for stroke victims. Essentially it is AI-charged push notifications + a group chat. It’s deceptively simple…but that works. Average time to notify a specialist in standard of care was 58.72 minutes vs. 7.32 minutes with Viz.ai. In 2021 Viz.ai successfully replicated this workflow for pulmonary embolism. commercial launch of AI-driven solutions for acute pulmonary conditions. Diagnosis and care coordination of patients suffering from pulmonary embolism (PE) can be challenging, with the average arrival-to-treatment times lasting more than 8 hours. Viz.ai uses deep learning to identify suspected pulmonary embolism disease in under two minutes. The Viz Platform is now utilized in over 850 hospitals across the U.S. and Europe. Obviously the approval from CMS has helped tremendously where Viz.ai demonstrated how this tech creates new workflow and measurably better outcomes for patients.

AI powered enrollment and more

Tempus released new product to identify potential therapeutic and clinical trial options for your patients. The platform claims to serve the customers in 3 keys ways:

• AI algorithms for molecular biomarker prediction: Their developing portfolio of digital pathology algorithms use a single whole slide H&E image to predict biomarkers, such as MSI or HRD status, for patients who are not ordinarily sequenced.
• AI algorithms for clinical trial enrichment: Their algorithms are being developed to use an H&E whole slide image to identify patients likely to contain characteristics relevant for clinical trial eligibility through our unique biomarker prediction technology.
• Solutions to digitize provider practice: They can serve as a partner to your practice in digitizing pathology workflow, from scanning to integrating algorithms into the existing workflow to uncover more potential options for patients.

Multi-modal model development 

On 20th May 2021 Google held its developer conference I/O and announced a new algorithm for their search engine: MUM, a Multitask Unified Model. For the two previous years, BERT was the underlying model for their search engine. BERT was a breathtaking release and was state-of-the-art until now, until MUM came. One of the most appealing aspect about MUM is the multi-modal data compatibility. This data point acts as a teaser of things to come making multimodal models part of mainstream.

Clinical decision-making in oncology involves multimodal data such as radiology scans, molecular profiling, histopathology slides, and clinical factors. Despite the importance of these modalities individually, no deep learning frameworks until last year has combined them all to predict patient prognosis.