focalinterest.net

AI Summary of Scholarly Research

This page presents an AI-generated summary of a published research paper. The original authors did not write or review this article. [See full disclosure ↓]

Deep learning distinguished three fibro-osseous jaw lesions

Written by

editor@focalinterest.com

in

Research area:medicine-clinicaloncology

What the study found

A multislide, weakly supervised deep learning model was the best-performing approach for distinguishing fibrous dysplasia, cemento-ossifying fibroma, and cemento-osseous dysplasia from histology slides. The model’s performance exceeded that of experienced oral pathologists when only histologic slides were used.

Why the authors say this matters

The authors say distinguishing these fibro-osseous lesions matters because they have different prognoses and require different clinical management. They conclude that the model could serve as a supportive tool alongside clinical, radiologic, and molecular data.

What the researchers tested

The researchers developed and validated a deep learning model using 1,218 hematoxylin and eosin whole slide images from 338 cases across 3 institutions. They compared 4 training strategies using a ResNet-50 backbone with loss functions and multiple-instance learning, including weakly and fully supervised models on single or multiple slides.

What worked and what didn't

In the test set, the weakly supervised multislide model performed best, with an area under the curve of 0.86 and accuracy of 0.71. The abstract says this model outperformed other models and exceeded the diagnostic accuracy of experienced oral pathologists, and heat maps suggested it identified key histomorphologic patterns relevant to the three diagnoses.

What to keep in mind

The authors note that the test cohort was limited in sample size and geographic diversity. They say more and more diverse cohorts would be needed to better support how generalizable the model is.

Key points

  • The model was trained on 1,218 hematoxylin and eosin whole slide images from 338 cases.
  • The best result came from a weakly supervised multislide approach.
  • That model reached an area under the curve of 0.86 and an accuracy of 0.71 in the test set.
  • The model outperformed experienced oral pathologists when only histologic slides were considered.
  • The authors describe limited sample size and limited geographic diversity in the test cohort.

Disclosure

Research title:
Deep learning distinguished three fibro-osseous jaw lesions
Authors:
A.B. Zhang, P.Y. Li, Jiang Xue, J H Zhang, Z. You, S H Ge, Z Xu, Z.P. Sun, D.X. Chang, L.S. Sun, T.J. Li
Institutions:
Beijing Jiaotong University, Beijing Jiaotong University, Chinese Academy of Medical Sciences & Peking Union Medical College, Chinese Academy of Medical Sciences & Peking Union Medical College, Chinese Academy of Medical Sciences & Peking Union Medical College, Chinese Academy of Medical Sciences & Peking Union Medical College, Chinese Academy of Medical Sciences & Peking Union Medical College, Chinese Academy of Medical Sciences & Peking Union Medical College, National Clinical Research, National Clinical Research, National Clinical Research, National Clinical Research, National Clinical Research, National Clinical Research, Peking University, Shandong First Medical University, Shandong Provincial Hospital, Shandong University, Shandong University
Publication date:
2026-02-24
DOI:
10.1177/00220345261415888
OpenAlex record:
View
AI provenance: This post was generated by gpt-5.4-mini (OpenAI). The original authors did not write or review this post.

More posts