番茄社区

A deep learning model was as effective as radiologists in detecting clinically significant prostate cancer on multiparametric MRI, a retrospective study suggested.

In an internal test set of 400 examinations, the model's performance did not differ from that of experienced radiologists in detecting clinically significant prostate cancer, with area under the receiver operating characteristic curves (AUCs) of 0.89 and 0.89 for the deep learning model and radiologists, respectively (P=0.88), reported Naoki Takahashi, MD, of the Mayo Clinic in Rochester, Minnesota, and colleagues.

Results were similar on an external test set of 204 examinations, with AUCs of 0.86 and 0.84 for the deep learning model and the radiologists (P=0.68), they noted in .

Moreover, Takahashi and colleagues found that the deep learning model plus radiologists performed better than radiologists alone in the external test set (AUC 0.89 vs 0.84, P<0.001).

"We believe that our model has the potential to assist radiologists in identifying clinically significant prostate cancer and facilitate lesion biopsy, hence improving the diagnosis of prostate cancer," they wrote.

The authors explained that machine learning or deep learning models are usually trained using ground truths, which include manually annotated regions of interest that are correlated with pathology results.

However, they pointed out that this is a resource-intensive and time-consuming process, "limiting the number of cases that can be used for model development."

Thus, they said their model is "unique" in that the ground truth labels contained only the presence or absence of clinically significant prostate cancer without requiring information about lesion location.

In an , Patricia M. Johnson, PhD, and Hersh Chandarana, MD, MBA, both of the NYU Grossman School of Medicine/NYU Langone Health in New York City, pointed out that while studies have consistently shown that MRI outperforms prostate-specific antigen (PSA) testing in screening for clinically significant prostate cancer, reducing the cost and increasing the accessibility of MRI "is critical for the viability of MRI screening programs."

This study "represents a step toward MRI screening for clinically significant prostate cancer by potentially decreasing the interpretation burden," they wrote.

Johnson and Chandarana noted that while the study used multiparametric MRI, "the field is shifting toward use of biparametric prostate MRI," which promises abbreviated examination times and minimizes patient discomfort.

With the "impressive results" demonstrated by deep learning with multiparametric MRI in the current study, extending this approach to biparametric MRI "could substantially amplify its utility and impact," they suggested, adding that training a similar model for biparametric MRI datasets "is a highly valuable direction for ongoing and future research."

Of note, Takahashi said in that he does not think the model can be used as a standalone tool, but should be used as "an adjunct in our decision-making process."

He added that he and his colleagues would like to conduct a prospective study examining how radiologists interact with the model's predictions. "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," he explained.

For this study, Takahashi and colleagues used data from patients without known prostate cancer at a single academic institution who underwent MRI from January 2017 to December 2019.

A total of 6,141 examinations from 5,555 individual patients met the inclusion criteria, with the final sample containing 5,735 examinations from 5,215 individual patients (mean age 66 years). Of these 1,514 examinations (1,454 patients) showed clinically significant prostate cancer.

Because the output from the deep learning model does not include tumor location, the researchers used a gradient-weighted class activation map (Grad-CAM) to show tumor localization. For true-positive examinations, Grad-CAM consistently highlighted clinically significant lesions, the authors said.

They also determined that integrating clinical features such as serum PSA levels and gland volumes with the deep learning image-only model and with radiologists' interpretations further improved the AUC on the internal test set.

Takahashi and colleagues acknowledged the study has limitations, including the fact that only radiologists who specialized in prostate MRI participated in the study.

"It is anticipated that the model will perform better than and further improve the diagnostic accuracy of trainees and general radiologists," they wrote.

Comment