Abstract:
Accurate 3D reconstruction of CT volumes from sparse and limited slices remains a core challenge in medical imaging, impacting both diagnostic confidence and data efficiency. This study investigates two deep learning approaches—self-supervised and supervised pipelines—for volumetric CT reconstruction from sparse input slices. In the self-supervised approach, pretraining is performed on the LIDC-IDRI public dataset (239 volumes, 40,690 slices) using contrastive and slice-order prediction losses, enabling strong feature learning from unlabeled data. Separately, a supervised Conv3D Autoencoder is fine-tuned using expert-annotated clinical volumes to maximize domain adaptation and reconstruction fidelity.Robust data preprocessing including denoising, augmentation, and volume alignment—ensures generalizable model input. Model performance is assessed on held-out test sets using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), with quantitative and qualitative comparisons against recent state-of- the-art modelssuch as XctDiff and diffusion-based approaches. Experimentalresults show that the supervised Conv3D Autoencoder achieves a mean PSNR of 38.98 dB and SSIM of 0.956, substantially outperforming all baselines and ensuring superior anatomical detail retention.Taken together, the findings demonstrate that while both self-supervised and supervised strategies are promising, the supervised Conv3D Autoencoder delivers the highest fidelity for sparse-view 3D CT reconstruction. This approach enables reliable, efficient, and low-dose CT imaging, with potential to support both retrospective 3D data recovery and future prospective low-dose clinical workflows.
