Deep learning for differential diagnosis of parotid tumors based on 2.5D magnetic resonance imaging

PMCID: PMC12180351

PMID: 40531801

DOI: 10.1080/07853890.2025.2520401

Journal: Annals of medicine

Publication Date: 2025-6-18

Authors: Mai W, Fan X, Zhang L, Li J, Chen L, et al.

Key Points

• 2.5D deep learning approach significantly improves diagnostic accuracy of parotid gland tumors
• Model achieved AUC of 0.86 and sensitivity of 0.78, compared to traditional model's 0.79 and 0.54
• Offers a non-invasive, potentially more reliable alternative to fine-needle aspiration for tumor diagnosis

Summary

This retrospective study addressed the critical challenge of accurately differentiating between benign and malignant parotid gland tumors (PGTs) using advanced deep learning (DL) techniques. By employing a novel 2.5D imaging approach that incorporates inter-slice information, the researchers developed a sophisticated diagnostic model that significantly outperformed traditional diagnostic methods. The study utilized magnetic resonance imaging (MRI) data from 122 patients, demonstrating that the deep learning model, particularly the 2.5D T2-weighted fat-suppressed (T2WI-FS) model, achieved superior diagnostic accuracy compared to conventional clinical approaches.

The research revealed that the deep learning model dramatically improved diagnostic performance, with an area under the curve (AUC) of 0.86 and sensitivity of 0.78, compared to the traditional model's AUC of 0.79 and sensitivity of 0.54. By integrating transfer learning and a self-attention mechanism, the model effectively addressed the challenges of diagnosing rare and complex salivary gland tumors, offering a promising alternative to invasive fine-needle aspiration (FNA) biopsies. The approach not only reduces patient discomfort but also provides a more consistent and automated method for tumor classification.