EasyLabels: weak labels for scene segmentation in laparoscopic videos

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Accurate segmentation of laparoscopic images is crucial for surgical guidance and analysis.
• Fully supervised methods require extensive pixel-level annotations, which are time-consuming and labor-intensive.

Purpose of the Study:

• To develop and validate a novel weak annotation approach for laparoscopic image segmentation.
• To demonstrate that this method can achieve accuracy comparable to fully supervised approaches with significantly reduced annotation effort.

Main Methods:

• Proposed a method using weak annotations in the form of stripes over objects.
• Employed partial cross-entropy as the loss function for a fully convolutional neural network.
• Generated dense pixel-level prediction maps for segmentation.

Main Results:

• Validated the method on three datasets, showing strong qualitative and quantitative results.
• Achieved at least [Formula: see text] of the accuracy of fully supervised methods.
• Required [Formula: see text][Formula: see text] less time for annotation creation compared to full supervision.

Conclusions:

• Laparoscopic data segmentation is feasible with minimal annotated data.
• The proposed weak annotation method maintains high accuracy comparable to full supervision.
• This approach offers a more efficient alternative for annotating laparoscopic images for segmentation tasks.