Automatic metabolic breast cancer staging using [¹⁸F]FDG PET/CT: comparison with nuclear medicine physician-based and clinical staging.

Abstract

Purpose: This study aimed to evaluate a deep-learning (DL)-based framework to automatically perform breast cancer (BC) metabolic staging on [¹⁸F]FDG PET/CT, and to assess agreement among DL-based, nuclear medicine (NM) physician-based, and clinical staging.

Methods: A total of 403 histologically confirmed BC patients who underwent whole-body staging [¹⁸F]FDG PET/CT were retrospectively included. All [¹⁸F]FDG avid lesions suspected of malignancy were segmented by an NM physician and classified into four key tumor regions: primary tumor (pT), regional axillary lymph nodes (ALN), extra-axillary locoregional nodes (extra-ALN), and distant metastases (dM). Data were split into training (n = 303) and testing (n = 100) sets using a stratified approach. Class-specific DL segmentation networks were developed. A post-processing pipeline was implemented to derive metabolic TNM staging from the DL-based segmentation. Using NM-based and clinical staging segmentation as reference, accuracy, sensitivity, and specificity were computed for T, N, and M. Segmentation performance was evaluated using the Dice coefficient (DC) and lesion detection (LD) metrics. NM-based metabolic staging was also compared with clinical staging.

Results: DL-based staging showed concordance with NM-based/clinical staging of 75/62%, 87/74%, and 83/82% for T, N, and M, respectively. For N3 (presence of extra-ALN) and M1 (presence of dM), where [¹⁸F]FDG PET/CT is particularly relevant, sensitivity/specificity of DL-based (NM-based reference) were 0.86/0.96, and 0.97/0.78, respectively. Segmentation performance was good to excellent for pT, ALN, and dM (median DC ≥ 0.83 and LD ≥ 0.78), and moderate for extra-ALN (median DC = 0.63 and LD ≥ 0.71). NM-based metabolic staging agreed with clinical staging in 63%, 80%, and 99% of cases for T, N, and M, respectively.

Conclusion: Although expert supervision remains essential, the developed DL-based framework demonstrates potential as a supportive tool for metabolic staging in BC patients, facilitating a workflow-efficient [¹⁸F]FDG PET/CT-based staging assessment.