Deep Learning-Based Intelligent Aided Diagnosis System for Ophthalmic Ultrasound Images

Objective: To design an intelligent assisted diagnosis system for ophthalmic ultrasound images based on a deep learning approach, which serves as an intelligent clinical diagnostic aid analysis tool for posterior ocular segment diseases. Methods: In order to accomplish multi-level feature extraction and fusion, two pre-trained network models-InceptionV3-Xception - are concatenated to create the InceptionV3-Xception fusion model. A classifier better suited for the task of multiple classification and recognition of ophthalmic ultrasound images is created in order to categorize 3402 ophthalmic ultrasound images. The model evaluation metrics include accuracy, macro-average precision, macro-average sensitivity, macro-average F1 value, subject working feature curves, and area under the curve. The model's credibility is evaluated by testing the model's decision basis using a gradient-weighted class activation mapping method. Complete developing the ophthalmic ultrasound image intelligent assisted diagnosis system's human-computer interface using the PyQt toolkit. Results: The Inception V3-Xception fusion model’ s accuracy, precision, sensitivity, and area under the subject working feature curve on the test set were 0.9673, 0.9521, 0.9528, and 0.9988, respectively. The model's decision basis aligned with ophthalmologists' clinical diagnosis rationale. The auxiliary diagnosis system is portable, reliable, and has an easy-to-use interface. It can recognize ocular ultrasound images intelligently and analyze thermograms in an interpretable manner. Conclusion: The deep learning-based intelligent aided diagnosis system for ocular ultrasound images, which can quickly and accurately identify six distinct types of ophthalmic ultrasound images, aids in the intelligent evolution of ophthalmic clinical diagnosis.