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

Area of Science:

• Oncology
• Biomedical Imaging
• Computational Biology

Background:

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

Approach:

• Compared state-of-the-art DL models (ResUNet, UNet++, FracTALResNet, SenFormer, CEECNet) for segmenting nuclei and nucleoli in tumor biopsy volumes.
• Investigated semi-supervised learning using Cross Pseudo Supervision with unlabeled data.
• Trained and evaluated models on sparse manual labels from three in-house datasets.

Key Points:

• Evaluated the performance of various DL architectures for 3D segmentation tasks.
• Demonstrated the effectiveness of semi-supervised learning in improving segmentation accuracy.
• Achieved improvements in 3D Dice scores across tested models.

Conclusions:

• More complex DL models offer relative gains in segmentation performance.
• Semi-supervised learning effectively utilizes unlabeled data to enhance segmentation.
• These advancements are critical for mitigating the manual segmentation bottleneck in cancer research.