Tissue segmentation is an important task for histological deep learning, however, successful segmentation of rare tissue is more difficult. Therefore model development choices, such as loss function, are more important for rare segmentation. This work systematically evaluates the impact of loss function and model configuration choice over multiple metrics of performance using a real-world sparse segmentation problem: muscularis tissue segmentation in pediatric ileal histological images. This work unexpectedly shows that the most common segmentation loss function, Dice loss, did not produce models with stronger precision than sensitivity. Rather the converse was true, models trained using Dice loss over-predicted the positive class whereas models trained using Dice loss of the negative masks were more conservative when predicting the positive class. The best segmentation results were obtained when trivial supplementary labels were included, such as the presence or absence of any tissue or collagen at the pixel level, particularly when the decoder was shared between masks. Threshold choice was impactful, as expected, especially for precision. Correlation between model performance and the share of the target tissue, muscularis, in the specimen was high, with a larger share of muscularis in the specimen leading to higher model performance across metrics.