CMVFT: A Multiscale Attention-Guided Framework for Enhanced Keratoconus Suspect Classification in Multiview Corneal...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Keratoconus is a progressive corneal disease requiring early detection for effective management.
• Accurate identification of suspect keratoconus cases is crucial for timely clinical intervention.
• Current diagnostic methods may face challenges in distinguishing early-stage or suspect cases.

Purpose of the Study:

• To develop and evaluate a multiview fusion framework for identifying suspect keratoconus.
• To facilitate early clinical intervention by improving the detection of early keratoconus signs.
• To leverage advanced AI techniques for enhanced corneal topography analysis.

Main Methods:

• A retrospective cross-sectional study involving 573 corneal topography maps (normal, suspect, keratoconus).
• Development of the corneal multiview fusion transformer (CMVFT) integrating features from seven corneal maps.
• Utilized ResNet-50 for single-view feature extraction and a multiscale attention module (MSAM) for refinement, coupled with a fusion Transformer for multiview integration.

Main Results:

• The CMVFT framework effectively distinguished suspect keratoconus cases within complex feature spaces.
• Ablation studies confirmed the critical role of MSAM and the fusion Transformer in robust multiview integration.
• The approach successfully compensated for representation gaps in small corneal topography datasets.

Conclusions:

• This study pioneers the application of Transformer-driven multiview fusion in corneal topography analysis.
• CMVFT demonstrates significant potential for identifying suspect keratoconus, supporting early intervention strategies.
• The framework's ability to address small-sample dataset limitations offers promising implications for clinical practice.