Weakly Supervised 3D Semantic Segmentation Using Cross-Image Consensus and Inter-Voxel Affinity Relations.

Area of Science:

• Medical imaging
• Computational biology
• Computer vision

Background:

• 3D semantic segmentation of volumetric images is crucial for analyzing complex biological structures.
• Existing methods often require extensive voxel-wise annotations, which are time-consuming and costly to obtain.
• Weakly supervised learning offers a promising alternative to reduce annotation burden.

Purpose of the Study:

• To introduce CIVA-Net, a novel weakly supervised method for 3D semantic segmentation of volumetric images.
• To demonstrate that image-level class labels are sufficient for training accurate segmentation models.
• To validate the performance of CIVA-Net on both simulated and real cryo-electron tomography (cryo-ET) datasets.

Main Methods:

• CIVA-Net utilizes cross-image co-occurrence patterns to generate integral regions.
• The model explores inter-voxel affinity relationships to refine segmentation boundaries.
• It employs a weakly supervised learning strategy, requiring only image-level class labels for training.

Main Results:

• CIVA-Net achieves accurate volumetric segmentation using only image-level annotations.
• The model demonstrates comparable performance to state-of-the-art methods trained with stronger supervision.
• Empirical validation on simulated and real cryo-ET datasets confirms the effectiveness of the approach.

Conclusions:

• Weakly supervised learning, specifically with image-level labels, is a viable and efficient approach for 3D semantic segmentation.
• CIVA-Net offers a significant advancement in reducing annotation requirements for volumetric image analysis.
• The method shows potential for broad application in fields requiring detailed 3D structural segmentation.