Automatic lesion segmentation in ⁶⁸Ga-PSMA PET/CT and ¹⁷⁷Lu-PSMA SPECT/CT: added value of PET-guided SPECT in a...

Manual segmentation of prostate cancer metastases on PSMA PET/CT and SPECT/CT is time-consuming and poorly scalable, particularly in highly metastatic patients. This study evaluated nnU-Net-based automatic segmentation models trained on PET and SPECT either separately or jointly and assessed whether PET-derived information can improve SPECT lesion segmentation. Seventy-three patients with metastatic castration-resistant prostate cancer treated with ¹⁷⁷Lu-PSMA were retrospectively included: 48 from the Henri Becquerel Cancer Center (HBCC) and 25 from Nantes University Hospital (NUH). For each patient, ⁶⁸Ga-PSMA PET/CT and ¹⁷⁷Lu-PSMA SPECT/CT were acquired before and during the first treatment cycle respectively. All images were manually segmented by four nuclear medicine physicians in consensus. Four nnU-Net models were trained: M1 (PET/CT only), M2 (SPECT/CT only), M3 (joint PET/CT+SPECT/CT, unimodal input at inference), and M4 (SPECT/CT with PET/CT segmentation as a priori input). Models were first trained and internally validated on HBCC data, then retrained on the full HBCC cohort and externally validated on NUH data.

For PET/CT segmentation, M1 and M3 achieved comparable performance. M1 reached DSCs of 0.83 ± 0.19 (internal) and 0.76 ± 0.22 (external), while M3 achieved 0.83 ± 0.16 (internal) and 0.77 ± 0.21 (external). For SPECT/CT, the PET-guided model M4 (DSC: 0.63 ± 0.24 internal; 0.78 ± 0.14 external; PPV: 0.65 ± 0.26 internal; 0.75 ± 0.23 external) provided the best results. Compared with the SPECT-only model M2 (DSC: 0.61 ± 0.26 internal; 0.70 ± 0.25 external; PPV: 0.63 ± 0.25 internal; 0.71 ± 0.24 external), M4 showed no statistically significant difference in internal validation (DSC p = 0.35), while being statistically significant in external validation (DSC p = 0.014).

The nnU-Net framework enables accurate lesion segmentation on both ⁶⁸Ga-PSMA PET/CT and ¹⁷⁷Lu-PSMA SPECT/CT. While PET-only and joint PET+SPECT models perform similarly on PET images, incorporating PET-derived segmentations as prior information tends to improve SPECT/CT lesion segmentation. This PET-guided SPECT segmentation strategy leverages the higher spatial resolution of PET and represents a key step towards fully automated extraction of volumetric and dosimetric biomarkers for personalized prostate cancer treatment.