Semantic segmentation neural nets play a vital role in building Clinical Intelligent Decision Support Systems because their accuracy in most cases represents a lower bound for the accuracy of the whole system. This paper introduces deep research in loss functions for 3D medical segmentation of lung nodules. For the last few years, the combination of Dice loss and Cross Entropy loss was the ’golden standard’ in many semantic segmentation tasks, including medical imaging. However, the latest research proposes new, more suitable loss functions for medical semantic segmentation—Boundary DoU and Focal losses and their modifications. This work proposes a modification of Boundary DoU loss - Boundary DoU++ as a better choice for the segmentation task in case of highly unbalanced classes. The results show that the proposed loss outperforms other popular losses in nearly all cases. Solo Boundary DoU loss is also evaluated and shows high performance. Additionally, research studies the effect of training and inference of segmentation systems on different data setups: big datasets with different scanners and specific datasets for a scanner with and without contrast. Results show that models trained on scans without contrast may outperform in-domain models in scans with contrast. All experiments are conducted on the lung nodules segmentation dataset because this task is highly challenging due to the unbalanced data and tiny foreground regions. Source code of Boundary DoU++ is available: https://github.com/VSydorskyy/boundary_dou_plus_plus .