In the context of psoriasis lesion segmentation, traditional methods in image segmentation, including region-based approaches like Mask R-CNN, have shown notable success. However, the problem of segmenting irregular and intricate biological structures such as psoriasis lesions presents significant challenges that cannot be effectively captured by standard bounding box approximations. This paper explores the potential of transforming psoriasis lesion segmentation into pixel labeling tasks, which promise greater efficiency and integration with modern image-to-image networks widely used across various domains. We investigate the limitations of convolutional-based architectures in generating dense pixel embeddings capable of distinguishing individual lesions. Through both theoretical analysis and empirical evidence, we propose a novel approach that leverages semi-convolutional operations. These modifications, which introduce spatially guided pixel embeddings, offer substantial improvements over traditional methods and demonstrate their applicability to the segmentation of complex biological forms. By drawing connections to advanced techniques such as Hough voting and bilateral kernels steered by convolutional networks, we show that these methods significantly enhance the segmentation accuracy of psoriasis lesions, outperforming conventional region-based models like Mask R-CNN in terms of precision and adaptability to irregular shapes.