search

lseventeen / FR-UNet

[JBHI2022] Full-Resolution Network and Dual-Threshold Iteration for Retinal Vessel and Coronary Angiograph Segmentation
MIT License
112 stars 24 forks source link

A issue about dataset size and train #5

Open zengjianjun123 opened 1 year ago

zengjianjun123 commented 1 year ago

Hello! I am very interested in your excellent work, but when I was actively training with the drive data, I found that the training data set was divided into 48 small pieces, but the test was input as the whole blood vessel picture. Why is this? Or how much does breaking up into 48 chunks help improve your model metrics? How do you train your model without using blocks? Thank you!

lseventeen commented 11 months ago

Regarding this issue, I had previously read an article from around 2018 in MICCAI that specifically analyzed the impact on performance of dividing images into patches compared to using the original images. Currently, almost all methods of vascular segmentation involve breaking down the images into patches, which is also an effective method of data augmentation. The choice of a size of 48 was based on some prior work. However, no experiments have been conducted on this particular hyperparameter. If you decide not to divide into patches, then during data processing, simply comment out the code that divides into patches.