From Surface to Viscera: 3D Estimation of Internal Anatomy from Body Surface Point Clouds

Salih Furkan Atici, Eytan Kats, Daniel Mensing, Mattias P Heinrich

Keywords: Point Cloud Processing, PointTransformer, Organ Localization, Generative Deep Learning

Abstract: Accurate pre-scan positioning in diagnostic imaging is essential for guiding acquisition and reducing manual calibration time, yet current automated approaches typically rely on dense volumetric representations that are not leveraging the geometric properties or sparsity of surface representations. In this work, we introduce a sparse, point-cloud–based framework for estimating patient-specific 3D locations and shapes of multiple internal organs directly from the body surface. Our method leverages a new dual-encoder PointTransformer architecture: one encoder processes a mean-shape point cloud comprising 20 anatomical structures, while a second encoder extracts features from the patient’s body-surface point cloud. A shared decoder then predicts a deformed shape estimating the hidden individual anatomy patient. This enables accurate organ localization without volumetric rasterization or autoencoder-style bottlenecks. Trained on the German National Cohort (NAKO) dataset, our model substantially outperforms volumetric convolutional autoencoder (CAE) baselines, achieving a mean Chamfer Distance less than 5 mm and markedly lower surface-distance errors. These results demonstrate that sparse geometric learning with deformable point-cloud priors offers an efficient and highly effective alternative improving over dense convolutional deep learning methods for automated imaging workflow optimization.

Primary Subject Area: Geometric Deep Learning

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 232