Age-stratified deep learning model for thyroid tumor classification: a multicenter diagnostic study

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Oncology Diagnostics

Background:

• Thyroid cancer incidence is rising, with age being a key survival predictor.
• Current diagnostic methods may lead to overdiagnosis and overtreatment due to low mortality rates.
• The diagnostic impact of age in thyroid nodule classification requires further investigation.

Purpose of the Study:

• To develop an age-stratified deep learning (DL) model, ASMCNet, for thyroid nodule classification.
• To evaluate the impact of age stratification on DL model accuracy.
• To explore ASMCNet's potential to enhance radiologists' diagnostic performance and reduce unnecessary biopsies.

Main Methods:

• Retrospective analysis of 10,391 ultrasound images from 5934 patients across three hospitals.
• Development and validation of an age-stratified deep learning model (ASMCNet).
• Comparison of ASMCNet's performance against non-age-stratified models and radiologists using the DeLong test.

Main Results:

• ASMCNet achieved an AUROC of 0.906, sensitivity of 86.1%, and specificity of 85.1%.
• ASMCNet significantly outperformed non-age-stratified models (AUROC 0.867) and all radiologists.
• Radiologist performance improved with AI assistance, particularly using explaining heatmaps.

Conclusions:

• Age stratification in DL models significantly enhances thyroid tumor classification accuracy.
• ASMCNet demonstrates clinical applicability, assisting radiologists in improving diagnostic accuracy.
• The age-stratified approach is crucial for accurate thyroid nodule diagnosis, potentially reducing overtreatment.