Effect of Input Resolution on Retinal Vessel Segmentation Performance: An Empirical Study Across Five Datasets

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Deep learning models for retinal vessel segmentation often resize high-resolution fundus images.
• This resizing is done to manage GPU memory and standardize batch processing.
• The impact of image resizing on the detection of thin vessels is not well understood.

Purpose of the Study:

• To investigate how downsampling fundus images affects the detection of thin retinal vessels.
• To evaluate the adequacy of standard segmentation metrics like the Dice score for assessing thin vessel segmentation.
• To introduce and utilize a new metric for evaluating vessel segmentation stratified by vessel width.

Main Methods:

• Trained a UNet model on five diverse fundus datasets (DRIVE, STARE, CHASE_DB1, HRF, FIVES) at various downsampling ratios.
• Introduced a width-stratified sensitivity metric using Euclidean distance transform to assess thin, medium, and thick vessel detection separately.
• Kept all other training parameters constant across different downsampling ratios.

Main Results:

• For high-resolution datasets (HRF, FIVES), thin vessel sensitivity improved with downsampling, peaking at processed widths of 256-876 pixels.
• For lower-resolution datasets (DRIVE, STARE, CHASE_DB1), sensitivity was highest at native resolution and decreased with downsampling.
• Aggressive downsampling reduced thin vessel sensitivity by up to 15.8% in the DRIVE dataset, while Dice scores remained stable.

Conclusions:

• The optimal image resolution for deep learning-based retinal vessel segmentation is dataset-dependent.
• Standard metrics like Dice score are insufficient for evaluating the segmentation of thin vessels.
• A width-stratified approach is crucial for accurately assessing microvascular segmentation performance.