Deep learning models for benign and malign ocular tumor growth estimation.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Clinicians struggle to select appropriate image processing algorithms for medical imaging data.
• Deep learning models offer potential for analyzing complex medical images like OCT and OCT-A.
• Abundant medical imaging data supports the development of neural network-based methods.

Purpose of the Study:

• To present a strategy for selecting optimal deep learning models for medical image analysis.
• To evaluate deep learning variants for tumor region differentiation and 3D ocular tumor segmentation.
• To provide guidance for clinicians on model selection based on data characteristics.

Main Methods:

• Four deep learning variants were tested on optical coherence tomography (OCT) and OCT-Angiography (OCT-A) images from 50 mice eyes.
• Sensitivity analysis was performed using eight performance indices to assess model accuracy, reliability, and speed.
• An empirical expression was derived to aid in model selection based on image quantity and data type.

Main Results:

• U-net with UVgg16 excelled in segmenting malignant tumors with treatment, while U-net with Inception was best for benign tumors.
• Model performance improved exponentially with an increasing number of training images.
• OCT-Angiography analysis revealed neovascularization as a driver of tumor volume, and treatment transformed aggressive tumors into cysts.

Conclusions:

• The study provides a framework for selecting deep learning models in biomedical image analysis.
• U-net architectures demonstrate efficacy in analyzing ocular tumor characteristics from OCT and OCT-A data.
• The findings recommend adopting a systematic approach for model selection before clinical application.