| Abstrakt: |
The paper analyzed tools for creating annotations of medical images in image segmentation tasks. The performance of the well-known tools Supervisely, CVAT, and Segments.ai is compared with the information technology proposed in the work, which uses the Language Segment-Anything model with relevant text prompts and an automatic validation mechanism, based on the EfficientNet-B2 classification model. The main objective of the study was to determine the optimal approach to the automation of the image annotation process to ensure maximum speed, maintaining expert accuracy. The results showed that usage of the Supervisely tool reduced the initial annotation time to 39.7 seconds, but required additional 59.5 seconds to adjust the masks. CVAT, with its semi-automated tools, produced masks in 64.8 seconds, but required another 85.1 seconds for adjustments. In comparison, Segments.ai required a full manual annotation, which took 130.2 seconds. At the same time, the developed information technology, which uses the Language Segment-Anything model with task-specific text prompts and an additional automatic validation mechanism, significantly reduced the time for creating annotations to about 29.6 seconds per image, and also reduced the time for manual correction to 45.4 seconds. The developed information technology demonstrated high speed and accuracy in creating pseudo-masks, confirmed by experimental results. The main advantages of this approach are the decrease of time, needed for manual correction and increase the efficiency of the medical image annotation process. This work points out to the significant potential of using automated methods to accelerate annotation in the field of computer vision, improving the speed of performing medical data analysis tasks while maintaining the desired quality [ABSTRACT FROM AUTHOR] |
