xtie97 / las-net

Longitudinally-aware Segmentation Network for interim-therapy PET segmentation in pediatric Hodgkin lymphoma

Longitudinally-aware Segmentation Network (LAS-Net) 📑

This repository contains the code for the paper Automatic Quantification of Serial PET/CT Images for Pediatric Hodgkin Lymphoma Patients Using a Longitudinally-Aware Segmentation Network (under review).

Overview 🔎

Summary: A longitudinally-aware segmentation network (LAS-Net) trained on multi-center clinical trial data achieved high performance in automatic quantification of PET metrics for baseline and interim-therapy scans in pediatric Hodgkin lymphoma.

Key Points:

• 📈 LAS-Net leverages information from baseline PET images to inform and improve the analysis of interim PET, achieving an F1 score of 0.606 in detecting residual lymphoma, which was significantly better than models without longitudinal awareness (P<0.01).
• 🏆 When analyzing baseline PET/CT images, LAS-Net attained a mean Dice score of 0.772, demonstrating comparable performance to the best comparator method (P=0.32).
• 🏅 The quantitative PET metrics measured by LAS-Net, including qPET, ∆SUVmax, metabolic tumor volume (MTV) and total lesion glycolysis (TLG), were highly correlated with physician measurements, with Spearman’s correlations of 0.78, 0.80, 0.93 and 0.96, respectively.

Usage 🚀

We already released our model weights in Dropbox.

Citation 📚