Lesion segmentation on ¹⁸F-fluciclovine PET/CT images using deep learning

Area of Science:

• Radiology and Nuclear Medicine
• Artificial Intelligence in Medical Imaging
• Oncology

Background:

• 18F-fluciclovine PET/CT imaging is crucial for guiding post-prostatectomy salvage radiotherapy in prostate cancer.
• Accurate lesion detection and segmentation are vital for effective treatment planning.

Purpose of the Study:

• To evaluate the feasibility of deep learning methods for automated lesion detection and segmentation on 18F-fluciclovine PET/CT images.
• To compare the performance of different neural network architectures and input modalities (PET/CT vs. PET only).

Main Methods:

• Retrospective analysis of 84 prostate cancer patients from the EMPIRE-1 trial.
• Training and testing of three neural networks (U-net, Cascaded U-net, cascaded detection segmentation network) using a fivefold cross-validation and hold-out test.
• Quantitative evaluation using Dice Similarity Coefficient (DSC), Hausdorff distance (HD95), center-of-mass distance (CMD), and volume difference (VD).

Main Results:

• Deep learning models successfully detected 144/155 lesions (PET+CT) and 153/155 lesions (PET only).
• The best-performing network achieved an average DSC of 0.68 ± 0.15 and HD95 of 4 ± 2 mm.
• PET-only input demonstrated comparable or superior performance to PET+CT, with CT addition leading to more failures.

Conclusions:

• Deep learning methods are feasible for automated lesion segmentation on 18F-fluciclovine PET/CT images.
• This approach has significant potential to aid physicians in lesion identification and reduce contouring time.
• Automated segmentation can serve as a valuable second check in clinical workflows.