The Judicial Threshold for Deep Learning Patents

FEDERAL CIRCUIT ANALYSIS

Where the courts now draw the line between a patentable AI invention and an unpatentable application of an off-the-shelf model — and what it means for medical-device innovators.

Focus: Section 101 · AI/ML Claims    |    Case: Dental Monitoring v. Align    |    CAFC · July 7, 2026

For teams innovating at the intersection of artificial intelligence and medical devices, a recent Federal Circuit decision sharpens a threshold question every patent strategy must now answer: does your claim protect a genuine technical advance in the model, or merely point an existing neural network at a new problem?

On July 7, 2026, in Dental Monitoring SAS v. Align Technology, Inc., the U.S. Court of Appeals for the Federal Circuit affirmed a district court ruling that patent claims covering a deep-learning-based dental image analysis system are ineligible under Section 101. For anyone building AI into orthodontic, dental, or broader medtech products, the reasoning deserves close attention.

PATENTS AT ISSUE U.S. 11,049,248 & U.S. 10,755,409	SUBJECT MATTER Deep-learning dental arch image analysis
FRAMEWORK APPLIED Alice two-step (steps one & two)	OUTCOME Ineligible under § 101; SJ affirmed

01 — THE RULING

What the Court Decided

The dispute began when Dental Monitoring sued Align, alleging that Align's Invisalign virtual-care AI platform infringed patents directed to assessing the shape of an orthodontic aligner and analyzing dental arch images using a “deep learning device.” The district court structured the fight as a “patent showdown,” with each side selecting a representative claim for summary judgment.

Applying the two-step framework from Alice Corp. v. CLS Bank, the court held the claims ineligible. At step one, it placed the claims in the familiar ineligible category of collecting information, analyzing it, and displaying the results of that analysis. At step two, it found no inventive concept: the claims applied generic hardware — and a generic deep learning device — to carry out that abstract idea.

Dental Monitoring's central argument was that its trained device could quantitatively assess the separation between an aligner and a tooth with more precision than was previously possible, and that this was a specific technological solution. The Federal Circuit rejected the argument for a reason that patent drafters should underline: the claim language itself did not require any degree of quantitative precision beyond what an orthodontist could already achieve. The alleged improvement lived in the argument, not in the claim.

A “deep learning device” being trained on a specific subset of data is incident to the very nature of machine learning.

— U.S. Court of Appeals for the Federal Circuit

That single sentence carries the weight of the decision. The court held that training a model on a dataset — here, more than 1,000 images — is not itself a technological improvement, because training on data is simply what machine learning does. This reasoning extends the Federal Circuit's earlier decision in Recentive Analytics, Inc. v. Fox Corp., which held that applying generic machine learning tools to a new field of use — even one previously performed by humans — does not confer eligibility.

The court also closed a common escape route. Dental Monitoring argued that using deep learning for orthodontic guidance was not conventional when the patents issued. Citing BSG Tech LLC v. BuySeasons, the court explained that the step-two question is whether the claims contain an inventive concept, not whether the invention as a whole was unconventional at the time. The patents' own specifications undermined the case: they confirmed the deep learning device could be selected from a preset list of widely available neural networks. Novelty of application is not the same as an inventive concept in the technology.

It is worth noting for the record that this is a nonprecedential opinion. It does not bind future panels the way a precedential decision would. But it is a clear signal of how the current court reads AI-related claims, and it aligns directly with the precedential reasoning in Recentive. Treating it as anything less than a strategic warning would be a mistake.

02 — THE STRATEGY

What This Means for Your IP

THE CORE TAKEAWAY

The generic application of an off-the-shelf neural network to a new field — however novel that field — remains highly vulnerable to invalidation. Eligibility now turns on a claimed technical improvement to the AI model itself.

For medical-device innovators, and for the attorneys who protect their portfolios, the practical consequence is a shift in where the inventive weight of a claim has to sit. Selecting a known architecture, feeding it domain data, and applying it to a clinical task is increasingly treated as the expected use of the tool, not an invention over it. The advance has to be in the how, and it has to be in the claim.

Where eligible subject matter tends to live

▢ Architectural innovation. A specific, non-generic modification to the network structure — layers, connections, or processing pipeline — that solves a concrete technical problem.

▢ Training methodology. A novel training technique, loss function, data-representation scheme, or optimization approach — rather than the mere fact that training occurred on a dataset.

▢ Technical integration. A claimed improvement in how the model interacts with hardware, sensors, or a physical process, producing a measurable technical effect the claim actually recites.

▢ Quantified, claimed effects. Improvements written into the claim language with specificity — not asserted in briefing after the fact — so the technical solution is on the face of the claim.

The distinction the court is enforcing is between a claim that reads “apply a deep learning device to analyze dental images” and a claim that recites a specific, described technical mechanism by which the model does something a generic model could not. The specification matters enormously here: robust technical-improvement language must be present throughout the specification, with a clear nexus to specific claim limitations that can serve as a technical hook for examiners and courts alike.

03 — THE RECOMMENDATION

A Proactive Portfolio Review

We recommend that clients with AI and deep learning assets — particularly in medical imaging, diagnostics, and connected-device platforms — treat this as a prompt to review their portfolios before a challenge forces the question. Two exercises are worth doing now.

First, audit issued and pending claims for the vulnerability pattern. Any claim whose inventive weight rests on applying a known model to a new dataset or clinical context should be flagged. The current PTO environment can make issuance more achievable, but that is a different thing from a claim that survives litigation. An issued patent is not a valid patent.

Second, redirect the drafting emphasis for new filings toward the technical advances your team actually made to the model. If the real innovation is in a training method, a data representation, or an architectural change, that is where the claims and the specification should concentrate, with the technical effect recited rather than merely argued.