Accurate assessment of oestrogen receptor (ER) status is critical in guiding endocrine therapy for metastatic breast cancer. Biopsies are commonly used to determine ER status, yet they are susceptible to sampling errors and may not reflect the heterogeneity across lesions. Imaging alternatives such as computed tomography (CT) and 18F-fluorodeoxyglucose (FDG) PET/CT are employed to evaluate anatomical and metabolic features but fall short in indicating ER functionality. The introduction of 18F-fluoroestradiol (FES) PET/CT offers a promising method for detecting functional ERs throughout the body. Two automated tools—a lesion detection algorithm and a concordance analysis system—have been developed to support visual interpretation of 18F-FES PET/CT images, enhancing diagnostic efficiency and facilitating therapeutic decision-making.
AI Detection of ER-positive Lesions
A convolutional neural network using Retina U-Net architecture was trained to identify ER-positive lesions on 18F-FES PET/CT scans. The training involved 52 scans, employing fivefold cross-validation to optimise robustness. The model used dual-channel inputs from both PET and CT modalities to improve localisation and accuracy. Results demonstrated a median sensitivity of 62 percent across all lesions with no false positives per participant. Sensitivity increased significantly to 90 percent for lesions with high uptake, defined as a maximum standardised uptake value (SUVmax) above 1.5, and to 80 percent for lesions with volume exceeding 0.5 cm³. Lesions in the chest region showed higher detection rates than those in skeletal sites, likely reflecting the imaging challenges in bony structures and fewer examples available for model training in those regions.
Must Read: 18F-FDG vs 18F-FES PET for Staging Low-Grade ER-Positive Breast Cancer
The model outperformed a second nuclear medicine reader in detecting lesions with higher uptake, indicating potential to support clinicians by reducing variability and workload in cases with extensive disease burden. However, smaller or low-uptake lesions remained difficult to detect accurately, limiting sensitivity in some patients. These limitations could be mitigated by increasing the size and diversity of training datasets, especially including more examples of subtle lesion presentations. Additionally, high physiological tracer uptake in organs such as the liver introduced areas where detection was inherently difficult for both AI and human readers.
Stratification of FES-avid Metastases
To determine whether patients presented with FES-avid metastases, the detection algorithm was combined with a classification scheme based on SUVmax thresholds. Patients with at least one lesion above an SUVmax of 2.5 were labelled as having FES-avid metastases. Those with lesions between 1.5 and 2.5 were categorised as likely, while those with no lesions above 1.5 were considered negative. Using this framework, the AI tool achieved a sensitivity of 90 percent in detecting participants with some degree of FES-avid disease. Nonetheless, the specificity was lower at 55 percent due to several false positives caused by non-malignant tracer uptake.
Most patients categorised as negative had undergone prior resection of the primary tumour, contributing to low or absent signal. In cases where disease burden was low or lesions were few, a single undetected or misclassified lesion could change a patient’s status. Therefore, while the system shows promise for high-throughput stratification, further refinement is needed to reduce the risk of mislabelling, especially in borderline cases. Improved specificity may be achieved through refined thresholds or additional input features that differentiate pathological from physiological uptake.
Assessment of Concordance Across Imaging Modalities
An automated concordance analysis was applied to evaluate consistency between 18F-FES PET/CT and standard imaging methods—CT and FDG PET/CT. Conducted in a subset of 25 patients, the analysis used software to match lesion regions across modalities. In 68 percent of patients, over half of all lesions were visible on 18F-FES PET/CT. A comparable proportion was observed for standard imaging. However, disease heterogeneity was prevalent, with lesions present on one modality but absent on another in most cases.
Invasive lobular carcinoma, known for its low metabolic activity, posed a particular challenge for FDG PET/CT, often resulting in under-detection. In these cases, 18F-FES PET/CT detected more lesions, demonstrating its clinical value, particularly in patients whose disease may otherwise go unnoticed. Automated concordance also reduced the time burden of manual lesion matching and provided a consistent approach to evaluating discrepancies across modalities.
One limitation was that the analysis depended on manually contoured lesions, underscoring the need for future work to develop fully automated end-to-end systems. By integrating detection and concordance algorithms, it would be possible to automatically assess inter-lesion heterogeneity, a known prognostic factor for therapy response. The current toolset already represents a substantial step forward in enabling full-body assessments to become more feasible within routine clinical workflows.
The integration of artificial intelligence tools into 18F-FES PET/CT imaging offers a valuable advancement for managing ER-positive metastatic breast cancer. The lesion detection algorithm provides high sensitivity, especially for lesions with greater uptake or larger size and enables rapid patient-level stratification for endocrine therapy. The concordance analysis tool reveals cross-modality discrepancies and supports more informed treatment planning, particularly in heterogeneous or difficult-to-image disease. Together, these tools have the potential to enhance diagnostic precision, streamline interpretation and improve patient care. Further development and validation are required to optimise specificity, integrate both algorithms into a unified pipeline and ensure broad applicability across imaging centres and patient populations.
Source: Radiology: Imaging Cancer
Image Credit:iStock
Miller R, Battle M, Wangerin K et al. (2025) Evaluating Automated Tools for Lesion Detection on 18F Fluoroestradiol PET/CT Images and Assessment of Concordance with Standard-of-Care Imaging in Metastatic Breast Cancer. Radiology: Imaging Cancer, 7:3.
