AI for Good

Chapter 25Automated Segmentation of Prostate Cancer Metastases

—Yixi Xu

Executive Summary

Automated and accurate segmentation of lesions in radiological images of metastatic castration-resistant prostate cancer can facilitate personalized radiopharmaceutical therapy and advanced treatment response monitoring. Here, we sought to develop a convolutional neural networks-based framework for fully automated detection and segmentation of metastatic prostate cancer lesions in whole-body Positron Emission Tomography/Computed Tomography (PET/CT) images. We used 525 whole-body PET/CT images of patients with metastatic prostate cancer. Those scans used the [¹⁸F]DCFPyL radiotracer that targets the prostate-specific membrane antigen (PSMA). We trained U-Net (1)-based convolutional neural networks to identify lesions on paired axial PET/CT slices.

We used both batch-wise Dice loss and weighted batch-wise Dice loss approaches to minimize error, and we quantified lesion detection accuracy of each approach, with a particular emphasis on lesion size, intensity, and location. We used 418 images for model training, 30 for model validation, and 77 for model testing. Further, we allowed our model to take up to 12 neighboring axial slices to determine how incorporating a three-dimensional context influenced model performance. We then selected the optimal number of neighboring axial slices that maximized the detection rate on the 30 validation images and trained five neural networks with different ...