An artificial intelligence framework and its bias for brain tumor segmentation: A narrative review

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neuro-oncology Research

Background:

• Artificial intelligence (AI) is crucial for brain tumor detection and diagnosis.
• Existing reviews on brain tumor segmentation lack analysis of AI risk-of-bias (RoB) linked to specific AI architectures.
• A comprehensive review is needed to address RoB in AI for brain lesion segmentation (BLS) across diverse architectures and imaging data.

Purpose of the Study:

• To systematically review and link RoB with different AI architectures used in BLS.
• To categorize AI-based BLS studies based on architectural evolution.
• To provide recommendations for mitigating RoB in AI for medical diagnosis.

Main Methods:

• A PRISMA strategy was employed, analyzing 75 studies from PubMed, Scopus, and Google Scholar.
• Deep learning (DL) studies were classified into four architectural classes: CNN-based, ED-based, TL-based, and hybrid DL (HDL).
• Studies were evaluated based on 32 AI attributes, including architecture, modalities, and performance metrics, and ranked using a composite bias score.

Main Results:

• The performance ranking of architectures for BLS is: Transfer Learning (TL) > Encoder-Decoder (ED) > Convolutional Neural Network (CNN) > Hybrid DL (HDL).
• Encoder-Decoder (ED)-based models demonstrated the lowest AI bias in brain lesion segmentation tasks.
• A bias cutoff tool (AP(ai)Bias 1.0) was utilized to categorize studies into low-, moderate-, and high-bias groups.

Conclusions:

• TL architectures offer superior performance in BLS, whereas ED architectures present the lowest RoB.
• The study provides three primary and six secondary recommendations to reduce RoB in AI for BLS.
• Addressing RoB is critical for the reliable application of AI in medical diagnosis and brain tumor detection.