An AI model has edged closer to being clinically useful for detecting prostate cancer on prostate-specific membrane antigen (PSMA)-PET/CT images, a team in Sweden has reported.
By doubling the training dataset and refining the architecture of a model they previously developed, the group significantly improved the model’s performance, noted lead author Elin Trägårdh, MD, PhD, of Lund University in Malmö, and colleagues.
“We achieved a substantial improvement in the performance of our fully automated AI-based method for detecting and quantifying prostate tumor and suspected lymph node and bone metastases on F-18 PSMA-1007 PET/CT images,” the group wrote. The study was published August 20 in EJNMMI Physics.
The widespread adoption of PSMA-PET/CT has significantly increased the workload for nuclear medicine departments and nuclear medicine physicians, the group explained. AI algorithms, particularly convolutional neural networks (CNNs), have demonstrated progress in image-recognition tasks, and integrating them clinically could potentially increase efficiency and reduce errors, they suggested.
To that end, the group developed a CNN model in 2022 that performed with a sensitivity on par with nuclear medicine physicians for detecting and quantifying prostate cancer tumors and metastases. However, the model had a higher number of false positive lesions compared to nuclear medicine physicians, especially for the detection of suspected lymph node metastases, they noted.
To overcome that limitation, in this study, the researchers used a total of 1,064 patient F-18 PSMA-1007 PET/CT scans (approximately twice as many as they used previously) to train the model, of which 120 were used as a test set. Nuclear medicine physicians manually annotated suspected lesions on the images as ground truth for a comparison with the model’s performance. In addition, the researchers compared the model’s performance to their previous model.
The main results demonstrated that the sensitivity of the new model remained comparable to that of nuclear medicine physicians, while its positive predictive value significantly improved. For example, the false positive rate per patient for suspected lymph node metastases decreased from 2.85 to 1.08, the researchers reported.
| Sensitivity of an AI model for detecting tumors and metastases on prostate cancer PET/CT scans | |||
| Manual readings | Old AI model | New AI model | |
| Prostate tumor/recurrence | 82% | 66% | 85% |
| Lymph node metastases | 86% | 88% | 91% |
| Bone metastases | 70% | 71% | 61% |
“In this study, we further advanced a fully automatic AI-based method for detection and quantification of tumor and suspected metastases on F-18 PSMA-1007 PET/CT images,” the group wrote.
However, although the overall performance was better for the new AI method compared to the old, the sensitivities of the new AI method were worse for bone metastases compared to the old AI method and the human readers, the group noted.
Ultimately, further research is warranted, and to promote transparency and facilitate other studies, the researchers have made their AI model freely available to the scientific community at recomia.org.
“We encourage independent validation across diverse clinical settings and imaging protocols to assess generalizability and support future integration into clinical workflows,” the group concluded.
The full study is available here.