Segmentation of retinal blood vessels is essential in medical image analysis and plays a central role in eye disease diagnosis. Deep learning approaches have recently emerged, providing good results in recovering the retinal vascular tree. However, it comes at the expense of numerous training-annotated datasets, which are particularly difficult and costly to obtain in the medical domain, as only experts can provide reliable annotations. We propose to lift this drawback by introducing and comparing a novel supervised modular model using information from second-order derivatives of the input image enhancing thin structures, under extremely low annotations. Firstly, we improve the LIOT (Local Intensity Order Transformation) pre-processing procedure by adding the orthogonal direction of the blood vessels. Then, a new adjustable loss function is used to train a U-Net with only a few of the available annotations. It relies on both a second-order functional for recovering thin structures such as blood vessels and the binary cross-entropy. In this work, we investigate four variants of the proposed loss function. A second, cascaded U-Net is added to polish and denoise the results. We demonstrate through numerical experiments the soundness, modularity, and practicality of our pipeline on two publicly available datasets: DRIVE and CHASEBD1.