Reliable segmentation of live and dead knots in X-ray computed tomography images is crucial for automated wood grading and sawing yield optimisation in the timber industry. Segmentation in wet-state X-ray computed tomography images remains challenging due to low contrast, noise, and the small, fragmented nature of dead-knot volumes. Under a controlled setting with a knot prior derived from the ground-truth whole-knot mask, which can be assumed to be available in industrial evaluations, we benchmark representative modern semantic segmentation networks using a unified preprocessing, training, and evaluation protocol. We compare both region-overlap performance and boundary-localisation behaviour. To ensure consistent geometric assessment, we applied a lightweight, anatomy-informed post-processing to all predictions to stabilise the dead-knot boundary, and we complement region-overlap metrics with pith-referenced boundary-localisation metrics that quantify dead-knot boundary error. SegResNet achieves the strongest overall performance, including a live-knot Dice of 0.89 and a dead-knot Dice of 0.66, and also yields the most stable dead-knot boundary localisation. Overall, under the idealised setting where the knot area is constrained by ground truth, the residual encoder--decoder network (SegResNet) is particularly effective for dead-knot delineation on wet-state X-ray computed tomography slices.