Deep learning algorithms for automated assessment of total and cancerous prostate gland volume based on PET/CT

Objectives: Uptake of PET tracers in the prostate gland and the size of the gland and a cancerous part of it may serve as guidance for management of patients with prostate cancer. A fast automated analysis of PET/CT studies providing imaging biomarkers reflecting both pathophysiology and anatomy may potentially improve diagnostic accuracy and treatment efficacy. We aimed to develop and validate a deep learning-based method for analysis of PET/CT studies of the prostate gland.

Methods: The deep learning based method included a convolutional neural network (CNN) trained by means of manually segmented PET/CT studies. The prostate gland, urinary bladder and rectum were segmented in the CT scans and abnormal choline uptake in the prostate was segmented using the PET scans, by experienced radiologists and nuclear medicine physicians. A total of 150 PET/CT studies were used as training group. A CNN, taking both the PET and the CT scans as input, was used to delineate the prostate in the PET scans and a similar model was trained to delineate the prostate in the CT scans. Based on the prostate delineation in the PET scan the maximum standard uptake value (SUVmax), the volume of abnormal choline uptake in the prostate gland, the mean SUV of that volume and the total lesion uptake defined as the product between this volume and the mean SUV, were automatically calculated. Based on the prostate delineation in the CT scan the prostate gland volume was calculated. After training, the deep learning method was applied to a validation group comprising 52 prostate cancer patients, who had undergone 18F-choline PET/CT prior to radical prostatectomy. Forty-five of these patients were operated within three months of the PET/CT study. The weight of the prostatectomy specimens after removal of the seminal vesicles were compared to the corresponding prostate gland volumes measured automatically in the CT studies. The specimens were processed according to routine department procedures and in agreement with recommendations from the International Society of Urological Pathology. A nuclear medicine specialist recorded “manually” SUVmax in the prostate gland in all PET/CT studies for comparison with the corresponding automatically derived SUVmax scores. The automated analysis was performed in less than a minute for a PET/CT scan.

Results: SUVmax was on average 7.5 (range 2.7-15.5) and the manual and automated SUVmax measurements were the same in 46 of the 52 patients. In the remaining 6 cases, high activity in the urinary bladder or rectum was falsely defined as prostate uptake by the automated

Methods: The prostate weight was on average 50g (range 20g-109g) in the 45 patients with a prostatectomy less than 90 days after the PET/CT scan. The average volume of the prostate gland automatically measured from the CT studies was 51mL (range 27-75mL). The correlation between prostate weight and the automated prostate volume measurements was 0.60.

Conclusions: Our deep learning based method for fast automated analysis of the prostate gland in PET/CT studies showed good agreement with manually obtained measurements and pathology-derived weight of the prostatectomy specimens suggesting that this approach may become a promising adjunct to quantitative assessment of PET/CT studies in prostate cancer patients. The method is now open for validation in future prospective clinical trials.