The study focuses on accurate computer segmentation of polyps on colonoscopy images, which is important for the early detection of colorectal cancer, but is hampered by glare, movement, overlapping structures, variability in the appearance of polyps, and erroneous or inconsistent annotations. The authors present a new dynamic-Nu T-Loss (DNA-TLoss) loss function based on the Student's t-distribution, which is designed to be more robust to extreme values. They introduce three novelties: a learnable degree-of-freedom parameter ν for each image that is predicted using a simple NuPredictor network, spatially varying pixel weights λ for more sensitive error assessment, and multiscale loss computation at multiple spatial resolutions to capture both coarse and fine details. The method is implemented in the U‑Net network with the ResNet‑34 coder and tested on five public files (CVC‑300, CVC‑ClinicDB, ETIS‑LaribPolypDB, Kvasir and CVC‑ColonDB). DNA‑TLoss achieved the lowest Hausdorff distance on all datasets, averaging 14.6% less than the original T‑Loss, with up to 45.96% less on CVC‑300. It also had the lowest false positive rate in all five databases, with an improvement of up to 38.7% on CVC‑300 and 24.5% on Kvasir versus T‑Loss. The method also showed best-in-class calibration, with an expected calibration error of only 0.44% on the CVC‑300, and outperformed the other compared methods in four out of five ensembles.