A three dimensional reconstruction of vertebrae from CT Scans is necessary for surgical planning, medical teaching and diagnosis. Traditional methods are computationally expensive, while state-of-the art deep learning techniques lack in effective handling of sparse or incomplete data.We propose a deep learning pipeline composed of transfer learning, convolutional autoencoders, and cGANs for reconstructing 3D point clouds of vertebrae from sparse 2D CT slices.The feature extractor, based on MobileNet, takes 2D input images and their augmentations and feeds them to an encoder-decoder network, producing 3D point cloud predictions. The generation of point cloud further improved under a cGAN framework with novel geometric loss functions, one of which is the peculiar shape combining cosine similarity, MSE, and Koenderink Shape Index. The proposed approach offers a 95.4% cosine similarity between predicted and ground-truth point clouds with just 64 sparse CT slices. Qualitative visualization depict anatomical plausibility and robustness against rotational artifacts. This work demonstrates that transfer learning via conditional GANs may create 3D reconstructions of the vertebral body in CT data with minimal doses, less acquisition time, and complete preservation of its clinical value. The method involves CT-machine-specific modeling and sparse-view medical image reconstruction.