Transfer learning for anatomical structure segmentation in otorhinolaryngology microsurgery

Area of Science:

• Artificial intelligence in medical imaging
• Microsurgical landmark segmentation
• Machine learning for surgical assistance

Background:

• Reducing the annotation burden is a key challenge in artificial intelligence (AI) research.
• Accurate landmark identification is crucial for effective microsurgical procedures.

Purpose of the Study:

• To evaluate the effectiveness of multi-stage transfer learning (TL) for improving AI-based landmark segmentation in microsurgery.
• To assess the performance of Convolutional Neural Networks (CNNs) with reduced training data and across different surgical scenarios.

Main Methods:

• Construction of multiple datasets for landmark segmentation using 41,257 labeled images across 6 microsurgical scenarios.
• Training datasets using a multi-stage transfer learning (TL) methodology.
• Evaluation of CNN performance with varying training dataset sizes and direct application of learned weights across scenarios.

Main Results:

• Multi-stage TL enhanced segmentation performance significantly compared to baseline methods (mIOU 0.8869 vs. 0.6892).
• CNNs demonstrated robust performance (mIOU 0.8917) even when training data was reduced to 10% of the original size.
• Directly applying weights across different surgical scenarios yielded an mIOU of 0.6190 ± 0.0789.

Conclusions:

• Transfer learning (TL) improves AI model performance in datasets with reduced size and increased complexity.
• Data-based domain adaptation is feasible across different microsurgical fields using AI.
• AI models can achieve robust performance in microsurgical landmark segmentation.