Medical image segmentation (MIS) serves as a crucial foundation for clinical imaging diagnosis. However, due to the small contrast difference between target objects in medical image data and the small scale of labeled samples, it is still difficult to establish a high-performance medical segmentation model under a few labeled datasets. To address this issue, we introduce a medical image segmentation approach that is weakly supervised and built upon a latent diffusion model (LDM). Specifically, the method comprises two latent diffusion models. First, a LDM is trained to progressively generate preliminary labeled images from noise under the weak supervision of bounding box annotations. Subsequently, a classifier-free guidance (CFG) latent diffusion model is trained, which takes the bounding box and the original MRI slice images as input to guide the generation of pseudo-MRI images, and compares them with the preliminary labeled images generated by the first model to produce the final segmentation result. Experimental evaluations demonstrate the proposed approach's superior performance on the CHAOS dataset under data scarcity conditions, outperforming current weakly supervised segmentation benchmarks established by diffusion modeling techniques.