Sli2Vol: Annotate a 3D Volume from a Single Slice with Self-Supervised Learning (2105.12722v2)

Abstract: The objective of this work is to segment any arbitrary structures of interest (SOI) in 3D volumes by only annotating a single slice, (i.e. semi-automatic 3D segmentation). We show that high accuracy can be achieved by simply propagating the 2D slice segmentation with an affinity matrix between consecutive slices, which can be learnt in a self-supervised manner, namely slice reconstruction. Specifically, we compare the proposed framework, termed as Sli2Vol, with supervised approaches and two other unsupervised/ self-supervised slice registration approaches, on 8 public datasets (both CT and MRI scans), spanning 9 different SOIs. Without any parameter-tuning, the same model achieves superior performance with Dice scores (0-100 scale) of over 80 for most of the benchmarks, including the ones that are unseen during training. Our results show generalizability of the proposed approach across data from different machines and with different SOIs: a major use case of semi-automatic segmentation methods where fully supervised approaches would normally struggle. The source code will be made publicly available at https://github.com/pakheiyeung/Sli2Vol.

• The paper introduces Sli2Vol, which propagates a manually annotated 2D slice to an entire 3D volume using self-supervised learning, achieving impressive Dice scores over 80.
• The method learns an affinity matrix for adjacent slice reconstruction, ensuring consistent segmentation across different scanners and anatomical variations.
• This approach reduces dependence on extensive manual labeling, offering a scalable solution for automated medical imaging segmentation in clinical settings.

An Overview of Sli2Vol: 3D Volume Annotation from a Single Slice Using Self-Supervised Learning

The paper "Sli2Vol: Annotate a 3D Volume from a Single Slice with Self-Supervised Learning" introduces a novel approach to segmenting 3D medical volumes through the annotation of just a single 2D slice. The proposed method, Sli2Vol, leverages self-supervised learning to propagate segmentations across slices, offering a practical alternative to fully supervised methods which are often hampered by the high cost and domain-specific nature of annotations.

Methodology

Sli2Vol operates by learning an affinity matrix in a self-supervised manner to capture the correspondences between consecutive slices. This matrix plays a critical role in the propagation of a manually delineated mask from one slice to others within a volume. Notably, the approach is independent of the structure of interest (SOI) and maintains robustness across different scanner types and anatomical variations.

The training phase makes use of self-supervised techniques, where the model learns to reconstruct adjacent slices, allowing the learning of useful representations without requiring labeled data. This task is facilitated by the implementation of an edge profile generator which acts as an information bottleneck, preventing trivial solutions and encouraging the model to focus on structural features.

Experimental Results

The authors evaluated Sli2Vol on eight public datasets comprising various CT and MRI scans. Remarkably, Dice scores frequently exceed 80 (on a 0-100 scale) in cross-domain evaluations, underscoring the method's versatility and accuracy. This is particularly notable given that the model achieved these results without fine-tuning for individual datasets. Sli2Vol's agnosticism to the SOI or the specific imaging domain highlights its potential to generalize significantly better than traditional fully supervised approaches, especially in settings where domain shifts are prevalent.

Implications and Future Directions

The ability of Sli2Vol to generalize across datasets without specific tuning emphasizes its suitability for clinical environments where annotated data may be sparse or costly to accrue. By significantly reducing the labor associated with manual annotation, this method can facilitate broader applications of automated segmentation in medical imaging, across varied anatomical regions and imaging protocols.

From a theoretical standpoint, the paper contributes to the growing body of work in self-supervised learning for medical imaging. The proposed framework offers a scalable approach to segmenting complex structures, revealing opportunities for future research in enhancing network architectures and optimization techniques to further increase accuracy and efficiency.

Future developments could include integrating the method with interactive platforms, allowing clinicians to iteratively improve or adapt segmentations during analysis. Additionally, expanding on the verification module, potentially through more sophisticated machine learning techniques, may further reduce error accumulation and improve reliability.

In conclusion, Sli2Vol presents a comprehensive, self-supervised framework for 3D medical image segmentation, demonstrating robust cross-domain performance and reduced dependency on fully labeled datasets. This method holds promise not only for practical deployment in clinical settings but also as a foundational method for further advancements in automated medical image analysis.