DeepMorph: find complex structures before Raman confirmation

DeepMorph is an AI/deep-learning image-classification approach that enriches relevant particle candidates from a much larger population. Raman then provides chemical confirmation, particularly for rare API–excipient composite particles (AECPs).

Recover structures that simple morphology rules can miss

Conventional morphology targeting relies on fixed rules such as size, circularity, solidity or elongation. In complex suspensions, API, excipient and composite structures can overlap. Hard morphology filters may therefore exclude analytically relevant particles before Raman analysis.

DeepMorph uses image-trained models to retain complex candidate structures for Raman confirmation.

IMAGE MODEL → CANDIDATE ENRICHMENT → RAMAN CONFIRMATION

Classification performance on Raman-confirmed particle classes

Validated against Raman-confirmed particle classes. Correctly classified particles in this evaluation dataset:

~85% Cellulose · 233 / 273 correctly classified

~92% API · 266 / 288 correctly classified

~82% AECP · 232 / 282 correctly classified

Enrich rare candidates from a larger population

Application result: DeepMorph reduced a population of more than 10,000 particles to a focused candidate set for Raman confirmation.

10,256 particles in scan → 415 predicted AECP candidates → 143 Raman-confirmed AECPs

Confirmed AECPs represent 1.39% of the total scan population. The 143 confirmed candidates are an application result, not the classification accuracy of the image model.

Image-defined candidate → chemical confirmation

An authentic particle-level example illustrates the sequence from conventional morphology filters to candidate recovery and 2D Raman confirmation. This is separate from the classification-performance dataset above.

Authentic three-panel example showing conventional morphology targeting, DeepMorph candidate recovery and 2D Raman confirmation of an API–excipient composite particle.

Conventional morphology rules → DeepMorph candidate recovery → 2D Raman mapping. The same particle remains the analytical target.

A separate nasal-suspension validation using >30,000 fully Raman-mapped particles reported 95.0% overall particle-level classification accuracy. That dataset is presented on the OINDP page.