Efficient semi-supervised semantic segmentation of electron microscopy cancer images with sparse annotations.

Area of Science:

• Biomedical Imaging
• Computational Pathology
• Cancer Research

Background:

• Electron microscopy (EM) provides nanometer-resolution imaging crucial for understanding cancer therapy resistance.
• Manual analysis of EM data for structure identification is a significant bottleneck, requiring months per sample.
• Deep learning (DL) offers a promising solution to automate and expedite EM image analysis.

Purpose of the Study:

• To evaluate and compare state-of-the-art deep learning models for segmenting nuclei and nucleoli in 3D tumor biopsy volumes.
• To assess the effectiveness of semi-supervised learning (Cross Pseudo Supervision) using unlabeled data.
• To identify the most effective DL approaches for mitigating the manual segmentation bottleneck in EM data analysis.

Main Methods:

• Comparison of ResUNet, UNet++, FracTALResNet, SenFormer, and CEECNet models for segmentation.
• Implementation of semi-supervised learning (Cross Pseudo Supervision) with unlabeled data.
• Training and evaluation on sparse manual labels across three in-house annotated datasets.

Main Results:

• Demonstrated improvements in 3D Dice scores across evaluated deep learning models.
• Quantified the performance gains of more complex model architectures.
• Showcased the added value of semi-supervised learning in enhancing segmentation accuracy.

Conclusions:

• Advanced deep learning models offer substantial improvements over previous methods for EM image segmentation.
• Semi-supervised learning effectively leverages unlabeled data to boost segmentation performance.
• These findings pave the way for faster, more efficient analysis of EM data in cancer research.