Facilitating deep learning through preprocessing of optical coherence tomography images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning shows promise in ophthalmology but faces high implementation hurdles.
• Small-scale model training requires efficient methods to reduce computational burden and accelerate learning.

Purpose of the Study:

• To explore the impact of image preprocessing on deep learning model training efficiency in ophthalmology.
• To investigate if Fourier transformation and bandpass filtering can reduce computational demands for AI studies with limited data.

Main Methods:

• Optical coherence tomography images of various macular conditions were preprocessed using Fourier transformation and bandpass filtering.
• Generated high-frequency, low-frequency, and original image datasets were used to train identical deep learning models.
• Model performance and training speed (epochs) were compared across different image preprocessing techniques.

Main Results:

• Training with high-frequency images resulted in improved final performance and faster convergence (fewer epochs) compared to original images.
• Low-frequency image preprocessing did not yield meaningful performance improvements.
• The study demonstrates a significant acceleration in model training using specific image preprocessing.

Conclusions:

• Strategic image preprocessing, particularly high-frequency filtering, can accelerate deep learning model training in ophthalmology.
• This approach can facilitate the development of artificial intelligence models when faced with limitations in sample size or computational resources.
• Preprocessing offers a viable strategy to overcome common barriers in applying deep learning to ophthalmic image analysis.