JMIR Publications

ABSTRACT

Background:

Early-stage diagnosis of laryngeal cancer can significantly enhance patient survival rates and quality of life. However, the scarcity of specialists in low-resource settings constrains the timely review of flexible nasopharyngoscopy (FNS) videos, essential for accurately triaging at-risk patients.

Objective:

We introduce a preliminary AI-based screening framework to address this challenge in triaging at-risk patients in low-resource settings. The formative research tackles multiple specific challenges common in high-dimensional FNS data: first, the selection of clear, non-blurry images; second, the localization within frames that show an anatomical landmark of interest; and, lastly, the classification of patients into referral grades based on the recorded FNS frames.

Methods:

The system includes an image quality model (IQM) to identify high-quality endoscopic images fed into a disease classification model (DCM) trained on efficient GhostNet modules. To validate our approach, we curated a real-world dataset from 132 patients at a leading academic tertiary care center.

Results:

Based on this dataset, we demonstrated that the IQM quality frame selection achieved 0.895 for the area under the receiver operating characteristic curve (AUROC) and 0.878 for the area under the precision-recall curve (AUPRC). Given the high-quality images from the IQM, the DCM improved its performance by 38% in AUROC (from 0.600 to 0.833) and 8% in AUPRC (from 0.840 to 0.912). In addition, the AI model achieved a 2.5 times faster inference time than the ResNet50.

Conclusions:

This study demonstrated the feasibility of an AI-based screening framework designed for low-resource settings, demonstrating its capability to screen patients accurately and efficiently. This approach promises substantial benefits for healthcare accessibility and patient outcomes in regions with limited specialist care.