Performance of Deep Transfer Learning for Detecting Abnormal Fundus Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Non-mydriatic fundus photography is crucial for detecting retinal abnormalities.
• Accurate and efficient detection of these abnormalities is essential for timely diagnosis and treatment.
• Deep transfer learning (DTL) offers a promising approach for automated image analysis in healthcare.

Purpose of the Study:

• To develop and validate a DTL algorithm for detecting abnormalities in fundus images.
• To assess the performance of the DTL model using both internal validation and an external test dataset.

Main Methods:

• A DTL model was developed using the inception-ResNet-v2 architecture on 929 fundus images.
• Image preprocessing included normalization.
• The model was validated on an internal dataset (273 images) and tested on the Messidor dataset (273 images).

Main Results:

• Internal validation showed high performance: AUC 0.997%, sensitivity 97.41%, accuracy 97.07%, specificity 96.82%.
• Testing on the Messidor dataset yielded: AUC 0.926%, sensitivity 88.17%, accuracy 87.18%, specificity 86.67% for detecting abnormal fundus images.
• The DTL model demonstrated strong capabilities in classifying fundus images.

Conclusions:

• DTL exhibits high sensitivity and specificity for detecting fundus-related diseases.
• Further research is needed to refine the DTL method.
• Evaluating DTL's applicability in community healthcare settings is recommended.