According to a study published in Radiology, a deep learning algorithm performs as well as an abdominal radiologist in detecting clinically significant prostate cancer on MRI.
Takahashi. Image Credit: Radiological Society of North America
The researchers expect the model to help radiologists identify prostate cancer more effectively.
Globally, prostate cancer ranks as the second most frequent cancer among men. Radiologists commonly use multiparametric MRI to identify clinically significant prostate cancer. The Prostate Imaging-Reporting and Data System, version 2.1 (PI-RADS), represents the results. Lesion categorization by PI-RADS is not without restrictions, though.
The interpretation of prostate MRI is difficult. More experienced radiologists tend to have higher diagnostic performance.
Naoki Takahashi, MD, Study Senior Author and Radiologist, Mayo Clinic
Prostate MRI using AI algorithms has demonstrated the potential for enhancing cancer diagnosis and lowering observer variability. The necessity for a radiologist or pathologist to annotate the lesion during model creation and again during model re-evaluation and retraining following clinical application is a significant disadvantage of current AI techniques.
Takahashi added, “Radiologists annotate suspicious lesions at the time of interpretation, but these annotations are not routinely available, so when researchers develop a deep learning model, they have to redraw the outlines.”
Dr Takahashi and colleagues created a novel type of deep learning model that can predict the existence of clinically significant prostate cancer without requiring information about lesion location. They compared its performance to that of abdominal radiologists in a large cohort of individuals without known clinically significant prostate cancer who had MRI at multiple sites inside a single academic institution.
The researchers developed a convolutional neural network (CNN) to detect clinically significant prostate cancer using multiparametric MRI.
Improving Cancer Detection Rates with Fewer False Positives
Out of 5,215 individuals, 5,735 examinations revealed 1,514 cases of clinically significant prostate cancer. The deep learning model's clinically significant prostate cancer detection performance was not different from that of experienced abdominal radiologists on the 400 internal and 204 external test sets.
On both the internal and external test sets, the deep learning model and radiologists’ results outperformed radiologists alone.
The location of the tumors was determined by the researchers using a tool known as a gradient-weighted class activation map (Grad-CAM), as the deep learning model’s output does not include this information. According to the study, Grad-CAM reliably identified the clinically significant prostate cancer lesions for true positive examinations.
According to Dr Takahashi, the model has the potential to support radiologists by enhancing MRI diagnostic performance by detecting cancer at a higher rate while producing fewer false positives.
Dr Takahashi noted, “I do not think we can use this model as a standalone diagnostic tool. Instead, the model’s prediction can be used as an adjunct in our decision-making process.”
The dataset has grown over time and includes twice as many instances as it did in the initial investigation. The next stage is prospective research that looks at how radiologists engage with the model's prediction.
“We’d like to present the model’s output to radiologists and assess how they use it for interpretation and compare the combined performance of radiologist and model to the radiologist alone in predicting clinically significant prostate cancer,” Dr Takahashi concluded.
Journal Reference:
Cai, J. C., et. al. (2024) Fully Automated Deep Learning Model to Detect Clinically Significant Prostate Cancer at MRI. Radiology. doi:10.1148/radiol.232635