Abstract
In medical imaging, incomplete anatomical data often arises from limited fields of view, surgical interventions, or trauma. This paper introduces "Anatomy Completor," a novel 3D denoising auto-encoder (DAE) framework designed for the simultaneous reconstruction of multiple missing anatomical structures. By transitioning from direct mapping to a one-to-one residual learning paradigm and utilizing an aggregated loss function, the framework achieves high-fidelity 3D geometric reconstructions. Evaluated on the TotalSegmentator dataset, the proposed method demonstrates superior performance with a Mean Dice Similarity Coefficient (DSC) of 0.931, even when 40% of the anatomical volume is missing.
Introduction
Radiologists and surgeons frequently encounter incomplete 3D anatomical data. While 2D image inpainting is well-established, 3D anatomical shape completion remains challenging due to the complexity and variability of human anatomy. Most existing methods focus on single, simple structures like individual bones. However, clinical reality often involves multiple missing organs or tissues. Anatomy Completor addresses this by identifying and reconstructing multiple anatomies simultaneously within their natural geometric constraints.
Methodology
The framework is built upon a 3D Denoising Auto-Encoder (DAE). The authors contrast two training approaches: many-to-one mapping and one-to-one residual mapping. In the residual mapping paradigm, the network learns to predict only the missing components (the residual) rather than the entire volume. This approach significantly reduces false negatives and simplifies the learning task. To ensure robustness, a Loss Aggregation scheme is employed, where the model is trained on various incomplete versions of the same subject, with the Dice loss aggregated to encourage a generalized understanding of anatomical spatial relationships.
Experiments and Results
The researchers utilized the TotalSegmentator dataset, comprising 1,024 CT images, to train and validate the model. The residual-based model with aggregated loss ($DAE_{agg+res}$) outperformed baseline models across all quantitative and qualitative benchmarks. It achieved a Mean DSC of 0.931 on the most challenging test cases (40% volume removal). Qualitatively, the model produces anatomically plausible shapes for large organs like the liver and lungs, as well as complex structures like the aorta, ensuring they align correctly with the skeletal "container" provided by the ribs and spine.
Discussion and Conclusion
While highly effective, the model faces challenges with "terracing" artifacts due to voxel resolution and the reconstruction of very thin structures like individual ribs. Future work aims to expand the anatomical scope to include the skull and brain, and to utilize the framework for semi-supervised learning. In this context, the "Anatomy Completor" could generate pseudo-labels to significantly accelerate the manual annotation process for large-scale medical datasets.
Summary updated