Repeatability of two semi-automatic artificial intelligence approaches for tumor segmentation in PET

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Quantitative imaging biomarkers

Background:

• Positron emission tomography (PET) is crucial for cancer staging and treatment monitoring.
• Accurate segmentation of metabolic active tumor volume (MATV) is vital for prognostic evaluation and treatment efficacy assessment.
• Developing precise and repeatable segmentation algorithms for PET imaging remains a significant challenge.

Purpose of the Study:

• To compare the repeatability of two semi-automatic artificial intelligence (AI)-based segmentation methods against conventional approaches for PET tumor imaging.
• To evaluate the accuracy and repeatability of AI-driven textural feature (TF) and convolutional neural network (CNN) segmentation algorithms.
• To determine the suitability of AI-based methods for segmenting primary tumors and metastasis in lung cancer PET datasets.

Main Methods:

• Two AI-based semi-automatic segmentation methods (TF and CNN) were developed and compared with conventional methods (MV2, MV3, 41%SUVMAX, SUV4).
• Algorithms were trained, validated, and tested on a lung cancer PET dataset, with external validation on a retest dataset.
• Repeatability was assessed using test-retest coefficients (TRT%) and intraclass correlation coefficients (ICC), with accuracy measured by Jaccard coefficient (JC).

Main Results:

• Both AI-based methods demonstrated good segmentation accuracy (JC median TF: 0.7, CNN: 0.73).
• The AI approaches showed superior repeatability compared to most conventional methods, with lower TRT% (TF: 13.0%, CNN: 13.9%) and higher ICC (TF: 0.98, CNN: 0.99).
• The TF and CNN methods outperformed other conventional approaches in terms of both accuracy and repeatability.

Conclusions:

• Semi-automatic AI-based segmentation methods provide enhanced repeatability for PET tumor segmentation compared to conventional techniques.
• Both AI algorithms achieve accurate segmentation of primary tumors and metastasis.
• These AI-based approaches are promising tools for reliable PET tumor segmentation in clinical practice.