Lesion Segmentation Using Convolutional Neural Network for PET/CT-Guided Salvage Post-Prostatectomy Radiotherapy

Abstract

<h3>Purpose/Objective(s)</h3> Molecular imaging with novel PET radiotracers has been shown to significantly impact radiotherapy decision making, target definition, and cancer control in the setting of prostate cancer (PCa) recurrent after prostatectomy. We propose a deep learning-based method to automatically segment pelvic node and prostate bed lesions on 18F-fluciclovine (anti-1-amino-3-[18F] fluorocyclobutane-1-carboxylic acid, FACBC) PET/CT images for salvage post-prostatectomy radiotherapy. <h3>Materials/Methods</h3> Our proposed method, named hierarchical activation network, consists of three main subnetworks: a fully convolutional one-stage object detection (FCOS) network, a hierarchical module, and a mask segmentation network. FCOS is employed to detect the view-of-interests (VOIs) of prostate bed and pelvic nodal lesions. Hierarchical module is used to derive activation map to boost the classification accuracy around lesion boundary. The mask segmentation network utilizes the detected VOIs obtained from FCOS network and the activation map obtained from hierarchical module to perform binary segmentation of lesions within the detected VOIs. To evaluate the proposed method, we retrospectively investigated 146 lesions from 83 prostate cancer patients who had 18F-fluciclovine PET/CT scan acquired. Each dataset has lesions contoured by physicians based on 18F-fluciclovine PET/CT scan, which were served as ground truth and training targets. The proposed method was trained and evaluated by a five-fold cross validation strategy. Quantitatively, we characterized the accuracy of the proposed method by calculating 95 percentile of Hausdorff distance (HD₉₅), the distance between the centroids of the two contours, the absolute difference of volumes, and Dice similarity coefficient (DSC). <h3>Results</h3> The average HD₉₅, centroid distance, volume difference, and DSC among all lesions is 4.09±3.75 mm, 1.84±1.77 mm, 1.44±7.21 cc, and 0.72±0.17. There is no statistically significant difference observed between ground truth (manual contour) and our proposed lesion volumes on a patient-wise basis using a paired t-test of significance (p = 0.83). Quantitative evaluations show the accuracy of the proposed method and confirm the feasibility of the proposed segmentation method. <h3>Conclusion</h3> We demonstrated the accuracy of the proposed method by evaluating the discrepancy in the centroid and volume, as well as the overlaps with ground truth. It is shown that the proposed method has great potential in improving the efficiency and mitigating the observer-dependence in pelvic and prostate bed lesion contouring for radiation therapy, and facilitating the workflow for ¹⁸F-fluciclovine PET/CT-guided salvage post-prostatectomy radiotherapy.