PET segmentation models' overlap metrics fail to certify patient-level tumor burden

A new research paper published on arXiv explores the effectiveness of overlap metrics, such as Dice, in evaluating PET segmentation models for estimating metabolic tumor volume (MTV). The study found that while overlap metrics can rank models similarly to relative MTV error on average, they do not reliably certify patient-level tumor burden. Significant discrepancies were observed at the individual patient level, with models missing lesions or having opposing false-positive and false-negative volumes that masked spatial errors. The authors recommend that PET segmentation benchmarks include patient-level MTV error and missed-lesion counts alongside traditional overlap metrics. AI

IMPACT Highlights the need for more robust evaluation metrics in medical AI to ensure patient-level reliability.

RANK_REASON Research paper published on arXiv detailing limitations of evaluation metrics for AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Fluorodeoxyglucose Positron Emission Tomography/ Computed Tomography
• Hugging Face
• LesionTracer

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →