Statistical Shape Modeling (SSM) is a potent tool for analyzing anatomical differences across populations. It yields both population-wide and subject-specific shape statistics, offering valuable insights into anatomical variation. However, deep learning methods, while capable of extracting shape information directly from unsegmented images, face challenges such as the need for manual segmentations in training, as well as pre-processing requirements during both training and inference phases. Existing deep learning approaches also often focus solely on population-level shape statistics, neglecting subject-level shape statistics. We aim to address these limitations by mitigating the supervision bottleneck in training and inference phases for deep learning-based shape modeling methods. Firstly, we identify the need for improved automation in the localization of anatomies and consideration of rigid pose information. To this end, we introduce a novel deep learning framework tailored for anatomy localization and rigid alignment, which significantly enhances accuracy compared to previous approaches. Moreover, recognizing the importance of subject-level shape statistics and the challenge of handling multiple anatomies simultaneously, we present a novel framework enabling the prediction of both subject-level and population-level shape statistics for multiple anatomies within a single image. Our research highlights the significance of subject-level shape statistics in providing superior shape information, outperforming segmentation methods in medical imaging tasks. Despite advancements in automation during inference, the manual segmentation requirement during training remains a bottleneck. Therefore, we delve into investigating weakly supervised segmentation methods as potential alternatives. Through comprehensive qualitative and quantitative evaluations, we aim to determine whether these methods can alleviate the need for manual segmentation, thereby enhancing automation and efficiency in shape modeling pipelines.