Automatic Quantification of Serial PET/CT Images for Pediatric Hodgkin Lymphoma Using a Longitudinally Aware...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pediatric Hodgkin lymphoma requires accurate assessment of treatment response using serial PET/CT scans.
• Quantitative PET metrics are crucial for evaluating treatment efficacy but can be subject to variability.
• Longitudinal analysis of imaging data offers potential for improved accuracy in disease monitoring.

Purpose of the Study:

• To develop and validate a longitudinally aware segmentation network (LAS-Net) for quantifying serial PET/CT images in pediatric Hodgkin lymphoma.
• To assess LAS-Net's performance in lesion segmentation and quantitative PET metric extraction compared to physician measurements.
• To evaluate the model's ability to leverage longitudinal information for improved quantification.

Main Methods:

• Retrospective analysis of 297 pediatric patients with Hodgkin lymphoma using baseline (PET1) and interim (PET2) PET/CT scans.
• Development of LAS-Net, a deep learning model incorporating longitudinal cross-attention for analyzing serial imaging data.
• Evaluation of segmentation performance using Dice coefficients and lesion detection using F1 scores; quantification of PET metrics (MTV, TLG, qPET, ∆SUVmax) and comparison with physician-derived measurements using Spearman correlation.

Main Results:

• LAS-Net achieved a mean Dice score of 0.77 for baseline segmentation and an F1 score of 0.61 for residual lymphoma detection on interim scans, outperforming comparator methods.
• LAS-Net demonstrated strong correlations with physician measurements for quantitative PET metrics: qPET (ρ=0.78), ∆SUVmax (ρ=0.80), MTV (ρ=0.93), and TLG (ρ=0.96).
• High quantification performance was maintained in an external testing cohort, indicating generalizability.

Conclusions:

• LAS-Net significantly improves the quantification of PET metrics across serial scans in pediatric Hodgkin lymphoma.
• The model's longitudinal awareness enhances the analysis of multi-time-point imaging datasets for improved treatment evaluation.
• This deep learning approach offers a valuable tool for objective and accurate assessment in pediatric oncology imaging.