Attention to Lesion: Lesion-Aware Convolutional Neural Network for Retinal Optical Coherence Tomography Image...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate classification of retinal optical coherence tomography (OCT) images is crucial for diagnosing and grading macular diseases.
• Ophthalmologists rely on macular lesion morphology, size, and number for clinical diagnosis.

Purpose of the Study:

• To propose a novel lesion-aware convolutional neural network (LACNN) for enhanced retinal OCT image classification.
• To simulate ophthalmologists' diagnostic approach by focusing on lesion-specific regions.

Main Methods:

• A lesion detection network generates an attention map highlighting relevant regions in OCT images.
• This attention map guides a classification network, weighting local convolutional features.
• The LACNN integrates lesion information to improve classification accuracy and training efficiency.

Main Results:

• The LACNN method demonstrated effectiveness and efficiency in classifying retinal OCT images.
• Experiments were conducted on two independent, clinically acquired OCT datasets.
• The lesion-aware approach improved classification performance compared to standard methods.

Conclusions:

• The proposed LACNN method offers a promising approach for automated OCT image classification.
• By incorporating lesion-specific attention, the model mimics clinical expertise, improving diagnostic support.
• This AI-driven tool can assist ophthalmologists in the accurate and efficient grading of macular diseases.