[RSNA 2025] After Novo’s Alzheimer’s setback, Neurophet bets imaging AI can show when drugs really work

CHICAGO, Ill. -- By Kim Ji-hye / Korea Biomedical correspondent -- Neurophet went into this year’s meeting of the Radiological Society of North America (RSNA) betting that the Alzheimer’s story around GLP-1 drugs could break either way.

What its co-CEO Kim Dong-hyeon was sure of, though, was that the next generation of dementia trials would rise or fall on how precisely they used imaging and biomarkers.

“In Alzheimer’s now, you cannot stop at one marker that looks a little better,” Kim said in an interview with Korea Biomedical Review at Neurophet’s booth in Chicago on Tuesday. If drugmakers keep treating “early Alzheimer’s” as a single bucket instead of distinct biological subgroups, he warned, “phase 3 becomes a very expensive gamble.”

A day later, Novo Nordisk’s gamble with oral semaglutide came up short.

Neurophet co-CEO Kim Dong-hyeon spoke with Korea Biomedical Review at the company’s RSNA 2025 booth at McCormick Place in Chicago on Tuesday about how AI imaging and ARIA monitoring will shape the next wave of Alzheimer’s drug trials after semaglutide’s phase 3 setback. (Credit: Korea Biomedical Review)

At the Clinical Trials on Alzheimer’s Disease (CTAD) conference in San Diego on Wednesday, full phase 3 data from the 3,808-patient EVOKE and EVOKE+ trials showed that two years of once-daily oral semaglutide 14 mg failed to beat placebo on the primary CDR-Sum of Boxes endpoint, a composite score of cognition and daily function, or on key secondary cognitive and functional measures.

The GLP-1 drug nudged Alzheimer’s-linked biomarkers and inflammation markers in the right direction, but none of that translated into slower clinical decline, leading Novo to scrap a planned one-year extension.

For Kim, that biomarker–clinical disconnect is precisely the gap Neurophet is trying to fill.

“Drug developers love a clean biomarker plot,” he said. “But if imaging and fluid markers are not built into the trial from day one, in a way that actually defines who should be in the study, you end up proving a mechanism and failing a trial.”

Building an Alzheimer’s 'imaging spine'

Neurophet has spent the past decade building what Kim describes as an “imaging spine” for Alzheimer’s care and trials.

Its AQUA software quantifies regional atrophy and lesions on structural MRI, turning routine scans into what he calls “a deeper measurement tool” for brain reserve and small-vessel disease. SCALE PET does the same for amyloid, tau and other tracers. On top sits AQUA AD, tuned to disease-modifying Alzheimer’s antibodies like Eisai and Biogen’s Leqembi (lecanemab) and Eli Lilly’s Kisunla (donanemab), handling both eligibility screening and safety monitoring of amyloid-related imaging abnormalities (ARIA).

In practice, Kim said, AQUA AD is designed to answer two questions before an infusion is ordered: whether there is clearly enough amyloid to justify antibody treatment, and whether the burden of microbleeds and white-matter disease keeps ARIA risk within an acceptable range.

During treatment, the software supports serial ARIA checks and incorporates basic clinical data such as cognitive scores and APOE4 genotype, so that safety decisions are not based on MRI alone.

“ARIA is not something you just eyeball when MRI volume suddenly doubles,” Kim said. “You need consistent rules, and you need them to match how the drug was actually tested.”

That logic extends back into development. In late-stage programs, Neurophet increasingly acts as an imaging core lab, taking MRI and PET from dozens of centers, harmonizing the data and running its AI to extract more stable measures of atrophy, plaque load and vascular damage.

“In a global trial you’re not only fighting noise from one scanner,” Kim said. “You’re fighting the variability of many hospitals, many protocols, many cultures. Our job is to make the imaging comparable and then ask: for this drug, in what pattern of pathology does it really move the needle, and where does it not?”

GLP-1s are not dead, but the bar just went up

EVOKE and EVOKE+ were built on a pile of promising signals: preclinical models, massive real-world datasets and post-hoc analyses suggesting semaglutide might cut Alzheimer’s risk in diabetics. The trials enrolled amyloid-positive patients with mild cognitive impairment or mild dementia, let them stay on standard Alzheimer’s meds and ran semaglutide out to two years.

By CTAD, the answer was clear. CDR-SB decline was essentially identical between drug and placebo in both studies, and secondary measures told the same story.

In a substudy of about 200 patients, spinal fluid tests showed roughly 10 percent reductions in several forms of tau and in proteins linked to brain inflammation, and levels of high-sensitivity C-reactive protein, a blood marker of systemic inflammation, also fell. Safety and tolerability looked like semaglutide’s metabolic profile: mainly gastrointestinal side effects and weight loss, with no worrying new safety signal.

Kim doesn’t see that as the death of GLP-1s in neurology, but he does see a pattern.

He draws a distinction between treating an indication label, such as “early Alzheimer’s” by standard criteria, and treating a defined biological profile built from imaging and fluid markers. Blood tests will matter, he thinks, but not on their own.

“Blood will be part of the answer, imaging will be part of the answer,” he said. “If you only look at one, you will keep getting cases where the biology moves and the function does not, or the other way around.”

That’s one reason Neurophet is expanding SCALE PET to more tau tracers and working with partners on integrated readouts that combine PET, MRI and plasma instead of treating each silo separately. Internally, the team is already watching how CTAD and next year’s AD/PD meeting dissect the EVOKE biomarker data.

“The question is not just ‘did the drug work,’” Kim said. “It is, ‘in whom did it move the right markers, and what did their brains look like before they ever swallowed the pill?’"