Semantic segmentation dataset authoring with simplified labels.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Surgical Scene Understanding

Background:

• Semantic segmentation of laparoscopic images is crucial for surgical scene understanding.
• Creating accurate ground truth labels for medical datasets is time-consuming and requires expert annotators.
• Existing methods to reduce dataset authoring time include weak labels, pseudolabels, and synthetic data.

Purpose of the Study:

• To address the challenges of expert annotation in medical dataset creation.
• To enable non-medical annotators to contribute to medical image annotation tasks.
• To facilitate the creation of large-scale datasets for semantic segmentation.

Main Methods:

• Proposing simplified, semantically weak labels to reduce the need for medical expertise.
• Simulating dataset authoring with mixed medical and non-medical annotators to assess accuracy impact.
• Demonstrating a formulation for multi-dataset training using simplified labels.

Main Results:

• Simplified labels are a viable approach for dataset authoring.
• Incorporating non-medical annotators improves dataset creation, with medical annotators yielding higher accuracy gains.
• Multi-dataset training, including with incompatible classes converted to simplified labels, enhances performance.

Conclusions:

• Simplified labels provide a framework for both dataset authoring and multi-dataset training.
• Non-medical annotators can effectively contribute to semantic segmentation dataset creation.
• Converting incompatible dataset labels to simplified formats enables effective multi-dataset training.