How Insilico PET® Is Closing the Gap Between CT and PET Imaging

When we founded RADiCAIT, the core question we were trying to answer was deceptively simple: if we already know what healthy and diseased tissue looks like on CT, and we know what the same conditions look like on PET, can a sufficiently powerful model learn to translate between the two?

The answer — after four years of research, thousands of paired scan cases, and multiple clinical validation studies — is yes. With an important qualification: the model doesn't just look good. It performs at a level that is clinically meaningful.

The Gap That Needed Closing

PET (Positron Emission Tomography) imaging reveals the metabolic activity of tissue. Because cancerous cells, inflamed tissue, and degenerating neurons all exhibit abnormal metabolic patterns, PET provides diagnostic information that structural imaging — CT or MRI — simply cannot. This is why oncologists order PET-CT for staging, why neurologists use FDG-PET for Alzheimer's assessment, and why cardiologists rely on myocardial perfusion PET for viability evaluation.

The problem is access. A PET scanner costs £2–4 million to procure and requires specialist physics expertise, controlled radiotracer facilities, and a nuclear medicine licence to operate. The UK has approximately 100 PET scanners for a population of 67 million. The United States has about 3,000 for 330 million people — and even there, significant geographic disparities exist. In lower-income countries, PET is effectively absent from clinical practice entirely.

"Over 100 million people are diagnosed each year with cardiological, neurological, or oncological conditions where PET imaging could meaningfully inform their diagnosis or treatment. The vast majority never receive that scan."

The imaging bottleneck isn't a marginal problem. It is one of the most consequential access gaps in modern medicine.

The Insilico PET® Approach

Our hypothesis was that the structural and textural features visible in a CT scan encode latent information about the functional metabolic state of tissue. Not completely — CT and PET are genuinely different imaging modalities — but enough that a model trained on a sufficiently large, high-quality paired dataset could learn to extract that signal reliably.

Insilico PET® is built on a conditional generative architecture. The model takes a volumetric CT series as input and produces a synthetic functional volume in standardised uptake value (SUV) units — the same quantitative metric used in true PET imaging. The training dataset comprises tens of thousands of same-day paired CT and PET acquisitions, collected under rigorous IRB-approved protocols across multiple institutions.

Several design decisions were critical to clinical performance:

• Volumetric, not slice-by-slice: The model processes full 3D volumes, preserving contextual information across anatomy that slice-level models miss.
• Multi-scale feature extraction: The encoder architecture captures both fine-grained voxel-level texture and coarser organ-level anatomical context simultaneously.
• Uncertainty quantification: Every output includes a calibrated confidence map, flagging regions where the model's prediction is less certain — crucial for clinical trustworthiness.
• DICOM-native I/O: Input and output are handled as standard DICOM series, ensuring immediate compatibility with existing radiology infrastructure.

What 94% Concordance Actually Means

Our headline concordance figure — 94% — requires careful interpretation. It is not a single number pulled from a single study. It represents the weighted mean diagnostic agreement between Insilico PET® output and ground-truth PET imaging across three separate validation studies in oncology, neurology, and cardiology, each assessed by blinded consultant radiologists.

More specifically, "concordance" in this context means that in 94 out of every 100 clinically relevant cases, a radiologist reading the Insilico PET® output would reach the same diagnostic conclusion as a radiologist reading a true PET scan. This includes:

• Correct identification of metabolically active lesions
• Correct staging in oncology (N stage, M stage)
• Correct hypometabolic pattern characterisation in neurodegenerative disease
• Correct assessment of myocardial viability in the cardiac cohort

Where discordance occurs, it tends to be in cases involving very small lesions (<8mm), unusual anatomical variants, or CT acquisitions with significant artefact — categories where even experienced radiologists reading true PET scans show meaningful inter-reader disagreement.

The Clinical Significance of Residual Disagreement

A 6% non-concordance rate is not zero. It would be dishonest to present it as such, and clinically irresponsible to deploy Insilico PET® without acknowledging it. Our approach to the 6% is threefold.

First, we provide uncertainty maps that flag the cases where the model has low confidence — allowing radiologists to apply additional scrutiny precisely where it is needed. In our validation studies, model confidence calibration was excellent: high-confidence outputs showed >98% concordance, while flagged low-confidence outputs drove most of the residual disagreement.

Second, we are actively expanding our training dataset to address the case categories where errors concentrate. Ongoing data partnerships with academic medical centres are specifically targeting under-represented populations and scan protocols.

Third, Insilico PET® is designed as a decision support tool, not a replacement for radiologist judgement. The output augments clinical decision-making; it does not substitute for it. Regulatory strategy, clinical deployment protocols, and our product design all reflect this principle.

Looking Forward

The 94% we achieved is not a ceiling — it is a starting point. As our training datasets grow, model architectures improve, and clinical feedback loops mature, concordance will continue to increase. We expect to exceed 96% across all three indications within the next 18 months as our next-generation architecture moves into validation.

More importantly, the direction of travel in the broader field validates our approach. The literature on CT-to-PET synthesis has expanded dramatically since 2022, and the convergence of evidence suggests that the fundamental hypothesis — that CT encodes meaningful functional signal — is not just correct, but more richly exploitable than we initially expected.

For the 100 million patients who need functional imaging and don't receive it, closing the gap between CT and PET isn't an incremental improvement. It is a step change in what medicine can offer them.