Discriminative kernel convolution network for multi-label ophthalmic disease detection on imbalanced fundus image...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Ophthalmic diseases like glaucoma, diabetic retinopathy, and cataracts are leading causes of global visual impairment.
• Retinal structure analysis and fundus examination are crucial for diagnosing eye conditions.
• The Ocular Disease Intelligent Recognition (ODIR-5K) dataset is a key resource for multi-label, multi-disease classification of fundus images.

Purpose of the Study:

• To develop an efficient deep learning model for accurate multi-label classification of ophthalmic diseases from fundus images.
• To explore discriminative region-wise features without increasing computational load.
• To address class imbalance issues in fundus image datasets.

Main Methods:

• Introduction of the Discriminative Kernel Convolution Network (DKCNet), incorporating an attention block and a Squeeze-and-Excitation (SE) block.
• The attention block generates discriminative feature attention maps, while the SE block enhances channel interdependencies.
• Utilizing an InceptionResnet backbone for improved performance on the ODIR-5K dataset, coupled with label splitting and sampling techniques for class imbalance.

Main Results:

• DKCNet achieved high performance on the ODIR-5K dataset, with an AUC of 96.08%, F1-score of 94.28%, and kappa score of 0.81.
• The model demonstrated robustness by performing well on unseen fundus image datasets, indicating reduced bias towards training data.
• The proposed method effectively handles class imbalance through strategic label splitting and sampling.

Conclusions:

• DKCNet offers a computationally efficient and accurate approach for multi-label, multi-disease classification of ophthalmic conditions using fundus images.
• The model's strong generalization capability across different datasets highlights its potential for real-world clinical applications.
• This research contributes to advancing automated diagnostic tools for eye diseases, aiding in early detection and management.