Detection of Local Prostate Cancer Recurrence from PET/CT Scans Using Deep Learning

Abstract

Background: Prostate cancer (PC) is a leading cause of cancer-related deaths in men worldwide. PSMA-directed positron emission tomography (PET) has shown promising results in detecting recurrent PC and metastasis, improving the accuracy of diagnosis and treatment planning. To evaluate an artificial intelligence (AI) model based on [¹⁸F]-prostate specific membrane antigen (PSMA)-1007 PET datasets for the detection of local recurrence in patients with prostate cancer. Methods: We retrospectively analyzed 1404 [¹⁸F]-PSMA-1007 PET/CTs from patients with histologically confirmed prostate cancer. Artificial neural networks were trained to recognize the presence of local recurrence based on the PET data. First, the hyperparameters were optimized for an initial model (model A). Subsequently, the bladder was localized using an already published model and a model (model B) was trained only on a 20 cm cube around the bladder. Finally, two separate models were trained on the same section depending on the prostatectomy status (model C (post-prostatectomy) and model D (non-operated)). Results: Model A achieved an accuracy of 56% on the validation data. By restricting the region to the area around the bladder, Model B achieved a validation accuracy of 71%. When validating the specialized models according to prostatectomy status, model C achieved an accuracy of 77% and model D an accuracy of 77%. All models achieved accuracies of almost 100% on the training data, indicating overfitting. Conclusions: For the presented task, 1404 examinations were insufficient to reach an accuracy of over 90% even when employing data augmentation, including additional metadata and performing automated hyperparameter optimization. The low F1-score and AUC values indicate that none of the presented models produce reliable results. However, we will facilitate future research and the development of better models by openly sharing our source code and all pre-trained models for transfer learning.

Keywords: DenseNet; PC; PET/CT; [18F]-PSMA; artificial learning; deep learning; positron emission tomography; prostate cancer.

Figure A1

Cropped NIfTI of pseudo ID 00045.

Figure A2

NIfTI examples that might show reasons why the model makes mistakes.

Figure A3

Various ROI modifications.

Figure A4

16 layers of an image with visible PET channel and masking of UB, colon and hip bones.

Figure A5

16 layers of an image with visible PET channel and masking of everything except the prostate.

Figure 1

PSMA-PET/CT: (A) maximum intensity projection; axial corresponding CT (B) and PET (C) slices; (D) fused PET and CT slices. Example of a 78-year-old patient with osseus and lymphnodal metastases, as well as a local recurrence in the prostate bed (indicated by the red arrow).

Figure 2

Result of TotalSegmentator [20] on three patients in Subfigures (A), (B), and (C). White arrows indicate the detected prostate gland in patients A and B, while the red rectangle indicates the absence of the prostate gland in patient C. The white asterisks (*) indicate that the urinary bladder was successfully detected in all patients. Color indicates the organ that was detected by TotalSegmentator.

Figure 3

Confusion matrices of models A–D on the validation set (left) and model D on the test set (right). Columns are labels, and rows are predictions. The numbers within the cells are additionally color coded with a gradient from light blue (small numbers) to dark blue (large numbers).