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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Assessing residual disease in pediatric Hodgkin lymphoma using serial PET/CT scans is challenging due to subtle changes.
• Existing methods struggle with the nuanced detection of minimal residual disease in interim scans.

Purpose of the Study:

• To develop a longitudinally-aware segmentation network (LAS-Net) for accurate quantification of serial PET/CT images in pediatric Hodgkin lymphoma.
• To improve the detection and quantification of residual disease in interim-therapy PET scans.

Main Methods:

• Retrospective analysis of 297 pediatric Hodgkin lymphoma patients' baseline (PET1) and interim (PET2) PET/CT scans from clinical trials.
• Development of LAS-Net incorporating longitudinal cross-attention to integrate information from serial scans.
• Evaluation using Dice coefficients for baseline segmentation and F1 scores for interim scan detection, alongside correlation analysis (Spearman's ρ) for quantitative metrics against physician measurements.

Main Results:

• LAS-Net achieved a detection F1 score of 0.606 for residual lymphoma in PET2, outperforming comparator methods (P<0.01).
• Mean Dice score for baseline PET1 segmentation was 0.772.
• LAS-Net's quantitative PET metrics (qPET, ΔSUVmax, MTV, TLG) showed strong correlations with physician measurements (Spearman's ρ: 0.78–0.96).
• High performance was maintained in an external testing cohort.

Conclusions:

• LAS-Net significantly improves the quantification of PET metrics across serial scans for pediatric Hodgkin lymphoma.
• The model's longitudinal awareness enhances the evaluation of multi-time-point imaging datasets.
• This approach offers a more precise method for assessing treatment response and residual disease.