Influence of inputs for bone lesion segmentation in longitudinal ¹⁸F-FDG PET/CT imaging studies

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Bone metastases are common in metastatic breast cancer, leading to complications and necessitating treatment response assessment.
• Current deep learning methods often analyze single time-point images, limiting longitudinal assessment of treatment efficacy.
• Evaluating treatment response in bone metastases requires accurate segmentation across multiple patient scans.

Purpose of the Study:

• To develop and compare deep learning strategies for segmenting bone lesions in longitudinal 18F-FDG PET/CT scans of metastatic breast cancer patients.
• To investigate the generalizability of deep learning models trained at different time points (baseline and follow-up).
• To introduce an automated PET Bone Index (PBI) for quantitative assessment of bone metastases burden and treatment response.

Main Methods:

• Four 3D U-Net based deep learning networks were trained using different strategies: baseline (BL) images only, follow-up (FU) images only, both BL and FU images, and FU images with registered BL images and segmentations as input.
• Longitudinal 18F-FDG PET/CT scans from 45 metastatic breast cancer patients were analyzed.
• The PET Bone Index (PBI) was computed from segmentations to quantify bone metastases burden and evaluate treatment response.

Main Results:

• The network incorporating BL images and segmentations as prior knowledge achieved the highest Dice score (0.62) on FU acquisitions.
• Lower SUV uptake in FU images compared to BL images may explain the under-performance of networks trained solely on BL or combined data.
• The difference in PBI between baseline and follow-up acquisitions demonstrated strong potential for treatment response evaluation, achieving an AUC of 0.86.

Conclusions:

• An automated deep learning method effectively segments bone metastases on longitudinal 18F-FDG PET/CT images for metastatic breast cancer.
• Incorporating prior knowledge from baseline scans significantly improves segmentation performance on follow-up scans.
• The automated PBI offers a quantitative and reliable tool for assessing bone metastases burden and evaluating treatment response in clinical practice.