Training deep learning models for 3D segmentation remains a demanding task, requiring strategies that are both computationally efficient and capable of generalizing well to unseen data. This study presents DeCode, a novel conditioning approach that leverages label-derived features to dynamically support the decoder during the reconstruction process — with the goal of improving both training efficiency and segmentation quality.
At the core of DeCode is the use of conditioning embeddings built from learned numerical representations of 3D label shape features. During training, these embeddings guide the network toward more robust segmentation. At inference time, when labels are unavailable, the model predicts the necessary conditioning information directly from the input, using a feed-forward network trained alongside the main model.
DeCode was evaluated on synthetic data and cone-beam computed tomography (CBCT) images of teeth, using three CBCT datasets — one public and two in-house. The results demonstrate that DeCode consistently outperforms unconditioned baseline models in generalization to new data, achieving higher accuracy at a lower computational cost. This work is the first to explore conditioning strategies specifically in the context of 3D segmentation, offering a more efficient way to make use of annotated training data. Code and pre-trained models are publicly available.
Autors: Tomasz Szczepański, Michal K. Grzeszczyk, Szymon Płotka, Arleta Adamowicz, Piotr Fudalej, Przemysław Korzeniowski, Tomasz Trzciński, Tomasz Sitek