Multi-view deep learning for automated lymphoma staging from 18F-FDG PET/CT: physician-level accuracy with high-throughput workflow

Lymphoma staging plays a pivotal role in guiding risk-adapted therapeutic strategies. While the Ann Arbor system is the standard for staging lymphoma, current manual PET/CT evaluations suffer from interobserver variability and time inefficiencies. Recent advancements in artificial intelligence (AI) offer opportunities to objectively automate this workflow. This study aimed to develop a deep-learning model based on multi-view 2D maximum intensity projection (MIP) images from 18F-FDG PET/CT to automate lymphoma staging and compare its performance with nuclear medicine physicians. We retrospectively analyzed baseline 18F-FDG PET/CT scans of 227 patients with lymphoma. The deep learning models were constructed with 2D MIP, three dimensional (3D) PET and 3D CT images for binary classication (limited vs. advanced stage). The models were evaluated by 5-fold cross-validation. Ground truth were the stage results of senior nuclear medicine physicians (5–10 years). Staging performance of different models was assessed by Receiver operating characteristic (ROC) curves. McNemar’s chi-squared test was used to compare the staging differences between the best-performing model and the nuclear medicine physicians with early-career nuclear medicine physicians (1–3 and 3–5 years, respectively). The developed AI model, LASM-mMIP, utilized multi-view 2D MIP images across coronal, sagittal, left anterior oblique (LAO), and right anterior oblique (RAO) perspectives. It achieved an accuracy of 93.0% and an AUC of 0.942 for staging, outperforming other AI models based on single-view MIP images and 3D volumetric PET/CT data. LASM-mMIP exhibited significantly higher staging accuracy compared to junior nuclear medicine physicians (1–3 years, 84.58%, P < 0.05) and similar accuracy to intermediate-experience physicians (3–5 years, 96.9%, P = 0.095). Moreover, the model reduced staging time by four orders of magnitude compared to physicians [(2.13 ± 0.31) seconds vs. (17918.90 ± 56310.53) seconds, P < 0.05]. Failure analysis revealed that most false negatives were due to occult lesions, while false positives arose from physiological hypermetabolism. The LASM-mMIP model demonstrated high accuracy and significant temporal efficiency in lymphoma staging, offering performance comparable to early-career nuclear medicine physicians. By overcoming interobserver variability and reducing manual interpretation time, this AI-driven tool could streamline workflows and enhance decision-making in clinical settings. Future work should expand patient cohorts, explore additional clinical predictors, and refine stage-specific classification to further personalize lymphoma management.