<p>Segmenting endometrial cancer from CT and PET scans is difficult for many clinicians due to low tissue contrast, irregular tumor shapes, complex multi-modal cancers and signals. Fixed nonlinear convolutional layer architectures and transformer models are also unable to scale. U-KAN is a simple spline-based architecture that uses Kolmogorov-Arnold networks for learnable functional mappings. While it can scale to some spline-based mappings, it also has O(C²) complexity, indicating that it is similar to the previously mentioned approaches in high-resolution multi-modal architectures. This paper describes a computationally lightweight architecture named BlockSeg, built from two structured functional modules. They are the Structured Group Transform (SGT) and the Spline Gate (SG). SGT uses a group-wise, spline-inspired transformation to partition the functional modules according to the computational burden of these mappings, and then applies these mappings to CT and PET imaging. BlockSeg was evaluated in two configurations, each with CT segmentation and early fusion in CT and PET, respectively, on the ECPC-IDS dataset, a dataset created specifically for endometrial cancer segmentation. In CT imaging only, the model performs segmentation with a Dice score of 0.86 and an IoU of 0.74. In configuration CT and PET early fusion, the model has a Dice score of 0.82 and an IoU of 0.71. U-KAN performs slightly better with a Dice score of 0.88, while BlockSeg has a non-discernible effect on early fusion CT and PET and can scale functional transformations modularly, thereby maintaining competitive performance. In configuration CT and PET early fusion, BlockSeg performs comparably to CT-only imaging, with no discernible fusion effect, consistent with an early channel concatenation effect. This early channel concatenation does not exploit connective complementary cross-modal information. Thus, there is no cross-fusion discernible information, and subsequently no effect in CT and PET early fusion.</p>