Biomedical image segmentation demands exceptional precision, surpassing that typically required for natural image processing. The U-Net neural network, renowned in biomedical applications, integrates multiple auto-encoder elements to generate comprehensive hypotheses for extensive image segmentation. This work introduces U-Net++ architecture, characterised by dense skip connections and a cascade of encoder–decoder modules tailored for tasks such as chest CT scan diagnosis, nuclei and polyp segmentation, and MRI-based heart analysis. Comparative assessments reveal that U-Net++ achieves superior Intersection over Union (IOU) metrics compared to conventional U-Net architecture. Efficient delineation of brain tumour extent is a critical clinical need. With the addition of selective kernels and squeeze and excitation residuals to the U-Net framework, the automated method presented in this study produced an impressive mean dice score of 0.992 for Late Gadolinium Enhancement (LGE) segmentation in left ventricle (LV), right ventricle (RV), and left ventricular myocardium (LVM). This research introduces the NN U-Net network paradigm to the medical domain by recognising the dependencies inherent in preprocessing, training, and inference with the original U-Net model. NN U-Net exhibits adaptability and robustness, yielding superior dice scores across all classes. This study explores U-Net networks and their various architectural adaptations, documenting contributions published in biomedical image analysis.