One model to use them all: training a segmentation model with complementary datasets.

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Accurate surgical scene understanding is vital for computer-assisted surgery.
• Machine learning models for scene segmentation require extensive annotated data, which is often scarce.
• Existing datasets typically offer partial annotations for surgical anatomy.

Purpose of the Study:

• To develop a method for combining multiple, partially annotated datasets for improved surgical scene segmentation.
• To overcome the limitation of data scarcity in training machine learning models for surgical applications.
• To enable the use of multiple, complementary datasets for enhanced model performance.

Main Methods:

• Leveraging mutual exclusive properties of complementary annotations to maximize information gain.
• Utilizing positive annotations of other classes as negative samples.
• Excluding background pixels from binary annotations to avoid incorrect positive predictions.

Main Results:

• A DeepLabV3 model trained on the combined Dresden Surgical Anatomy Dataset showed significant improvement.
• The proposed method increased the overall Dice Score by 4.4% compared to individual model ensembles.
• Reduced class confusion, with a 24% drop in misclassifications for stomach and colon segmentation.

Conclusions:

• The developed method effectively improves surgical scene segmentation by integrating multiple complementary datasets.
• This approach alleviates the need for large, fully annotated datasets, demonstrating feasibility for practical applications.
• Paves the way for leveraging existing, partially annotated datasets to build robust surgical scene understanding models.