Which computer vision model (or LLM) for segmentation?

Question

I'm new here. I have nearly 2,000 bee images with hand annotations of masks of the full body. I want to automatically generate "head", "thorax" and "abdomen" from this full mask annotation. I tried different tools for it. At first, I tried the segment anything model 2 (SAM 2), but it didn't work well since it tried to generate the whole body mask instead of the head (the mask of the head is ambiguous and has high overlap with the thorax without any edge between them). Because of the lack of edges, comparing two points for segmentation also didn't work (I tried both SAM and SAM 2). Then I tried CLIP and gave text prompts for separating "head" from the body. But it returned nothing! I also tried manually finding the smallest contours in the full segmentation which can be its head. But It didn't work as the head often has no separate contours. I tried a very simple approach, like "take the first 20% of the full body mask as the head", but I am also finding it hard to detect the orientation of the bee as well, since different images have different orientations. Please suggest me any algorithm or any AI tool which I can use to separate "head", "thorax" and "abdomen" from the bee full body segmentation. Note that sometimes the shape of the full body segmentation is very ambiguous, and it just looks like a cylinder, but the colors of different areas may help.