Multiparametric characterization of periodontal tissues for inflammation monitoring using quantitative ultrasound, texture analysis and machine learning

The National Institute of Dental and Craniofacial Research awarded the Regents of the University of Michigan $379,320 on September 11, 2025, under the Trans-NIH Research Support program (CFDA 93.310) to develop noninvasive quantitative ultrasound biomarkers and machine learning classifiers for detecting periodontal inflammation.

The project establishes ultrasound-derived quantitative parameters and develops a machine learning classifier to characterize inflammatory conditions in periodontal tissues using high-frequency ultrasonography. The research applies quantitative ultrasound analysis and texture analysis to tissue scan data from a pre-existing in vivo cohort of periodontal swine scans, with preliminary findings showing ultrasound speckle statistics parameters discriminating tissue types with greater than 93% accuracy relative to histological ground truth. The work addresses the clinical need for noninvasive, objective techniques to replace the current standard of bleeding on probing, which is invasive, subjective, and a late indicator of disease progression.

Work is performed at the University of Michigan in Ann Arbor, Michigan. The period of performance runs from the award date through March 31, 2027.

Abstract

Approximately 42% of adults in the U.S are affected by periodontal diseases. These diseases impose excessive pain and systematic health issues on patients which even at their earliest stage negatively impact patients' everyday quality of life. Patients can be characterized by periodontal diseases based on the presence of a spectrum of oral inflammatory conditions, from gingivitis (i.e. early reversible inflammation) to periodontitis (i.e. irreversible damages). The current standard method in clinics for assessment of these diseases is bleeding on probing (BOP). BOP is invasive, subjective, semi-quantitative and a late indicator of disease progression. Thus, there is an urgent clinical need for research on noninvasive quantitative techniques for detecting these diseases for an improved diagnosis. To address this need, the current study centers around high-frequency ultrasonography, a well-established diagnostic modality for various biological tissues but with limited applications in periodontology. Our hypothesis is that quantitative parameters from ultrasound scan data and ultrasound images can be employed to estimate the extent of periodontal inflammation. Here, we aim at establishing ultrasound-derived quantitative biomarkers and a machine learning (ML) classifier for improved characterization of the inflammatory condition in a pre-clinical, pre-existing cohort consisting of in vivo periodontal swine scans. Our preliminary findings on quantitative ultrasound (QUS) analysis of two periodontal tissue types (gingivae and alveolar mucosa) strongly suggest that ultrasound speckle statistics parameters have potential as a biomarker. They discriminated tissues (>93%) with respect to histology as ground truth. Here, our proposed study will include a comprehensive QUS investigation of tissue scan data (specific aim 1), texture analyses of ultrasound B-mode images (specific aim 2) and a multiparametric ML classifier for periodontal inflammation classification combining parameters from QUS and texture analysis (specific aim 3). In the first aim, we will investigate longitudinal periodontal ultrasound data (baseline and inflammation) from ultrasound speckle models including homodyned-K, burr and nakagami. We will also assess the effect of periodontal inflammation on ultrasound attenuation. In the second aim, we will conduct four types of texture analysis techniques on longitudinal ultrasound images (image processing) including statistical and wavelet features to quantify the complex underlying structure of periodontal tissues, as a more sophisticated approach than traditional B-mode echogenicity analysis. We will assess the potential of each parameter as an individual inflammation biomarker. In our third aim, we will quantitatively analyze histology images to find potential correlation between histology parameters and QUS or texture parameters. Ultimately, we will perform a multiparametric analysis by building a ML classifier based on combined QUS and texture parameters to classify periodontal inflammation. The proposed research will provide novel insights into diagnostic values of QUS parameters and texture features for diagnosing and monitoring periodontal inflammation.