In many situations, the project manager knows very clearly that Deep Learning is the optimal solution to a segmentation problem. Still, it is a long way until this vision is put into practice. Let’s take medical segmentation as an example: we often want to segment organs, bones or pathologies. The general framework is very clear: we want a ground truth, feed it into the training system and hope that results will be good. However, getting a valid ground truth for medical segmentation is far from being simple. First, who is going to do the annotation? An expert radiologist or laymen who are given basic training? Due to the complexity and diversity of the human body, even expert radiologists may find that judgement for borderline cases is not clear-cut, resulting in a wide intra- and inter- observer variability. Another important aspect is quantity: it would be better to have thousands of images, but in practice this is not always possible, especially when precise annotations are required. We also need to understand what we have in those images: when a dataset includes many unrelated items or complex cases with multiple pathologies, its quality may suffer as a result. The quality itself depends on the resolution of the images coming from the different modalities (MRI, CT, etc.) in all their types. The project manager needs to make sure that the data received is in line with what the project requires. When data is gathered from big databases, there are many DICOM files which need to be sorted out to find what is relevant and what is not. Ground truth might be difficult to obtain for other reasons as well: some organs are very complex, like the airways leading to the lungs, and many times radiologists lack the time and patience to do this tedious task through a large dataset of images. In that case, extra tools are needed to facilitate the definition of the ground truth. We conclude saying that the first part of the project needs a thoughtful planning in order to achieve the ground truth. Deep Learning: from plan to practice Management Project Management Tip 9 Computer Vision News RSIP Vision ’s CEO Ron Soferman has launched a series of lectures to provide a robust yet simple overview of how to ensure that computer vision projects respect goals, budget and deadlines. This month we learn about Deep Learning: from plan to practice . It’s another tip by RSIP Vision for Project Management in Computer Vision . Extra tools are needed to facilitate the definition of the ground truth. We want a ground truth, feed it into the training system and hope that results will be good.