Not without Context-A Multiple Methods Study on Evaluation and Correction of Automated Brain Tumor Segmentations by...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Neurosurgery and neurology

Background:

• Accurate brain tumor segmentation is crucial for glioblastoma patient management.
• Manual segmentation is time-consuming and varies between providers.
• Automated deep learning tools are needed to improve segmentation efficiency and consistency.

Purpose of the Study:

• To identify expert criteria for evaluating automated brain tumor segmentation quality.
• To understand the thought processes of experts when correcting segmentations.
• To inform the development of better AI-driven segmentation models.

Main Methods:

• Utilized questionnaires and semistructured interviews with neuro-oncologists and neuroradiologists.
• Collected data between August and December 2021.
• Analyzed qualitative data using a combined deductive and inductive approach.

Main Results:

• Physicians heavily rely on patient and clinical context for segmentation quality evaluation.
• The intended clinical application is paramount in determining segmentation quality criteria and editing decisions.
• Personal physician beliefs about AI capabilities and inclusion of uncertain areas impact segmentation quality perception.

Conclusions:

• Findings enable the design of improved expert-centered evaluation frameworks for brain tumor segmentation models.
• This research can inspire the development of specialized metrics for training and evaluating AI segmentation models.
• Understanding expert criteria is key to advancing automated brain tumor segmentation in clinical practice.