A new deep learning model for MR can detect prostate cancer as accurately as an abdominal radiologist, according to a study published yesterday in Radiology. It won't ever replace radiologists, but the researchers envision it as a tool to help them improve prostate cancer detection.

Despite prostate cancer being the second most common cancer in men globally, interpreting prostate MR images is a challenge. Dr. Naoki Takahashi, the study's senior author and radiologist at Mayo Clinic, told HCB News that false positives are common because benign lesions often mimic prostate cancer.

AI algorithms for prostate MR are nothing new, as other studies have shown that they improve cancer detection and reduce observer variability. However, a radiologist or pathologist has to annotate the lesion when the model is under development and after clinical implementation when it’s reevaluated and retrained.
Also, when the developers are preparing the data set, they have to correlate imaging findings with the pathology report. If there are multiple lesions, it becomes a time-consuming process and may not even be feasible.

Takahashi and his team were able to develop a new type of deep learning model that can detect prostate cancer on MR without requiring that annotation. They achieved that by training a convolutional neural network, which is a type of AI that can detect subtle patterns in images better than the human eye.

They also used a large number of cases to train the model, which allowed it to learn the features of cancer without needing to know the location of the tumor.

"The task is like finding Wally in a picture," explained Takahashi. "Let’s assume some pictures contain Wally, but some don’t. [If you] show the pictures and tell a person whether Wally is present in that picture or not and you repeat that 1,000 times, then the person will eventually learn what Wally looks like and be able to tell if he is present or not."

To test its effectiveness, the team used data from patients without known clinically significant prostate cancer (csPCa) who underwent MR between January 2017 and December 2019 at one of the multiple sites that belonged to the same academic institution. A total of 5,735 exams were used and 1,514 of those showed csPCa.

They found that there was no difference between the deep learning model and the radiologists when detecting prostate cancer on MR. They also discovered that combining the model with the radiologist's findings resulted in better outcomes than the radiologists alone.

After they completed the study, the team continued to increase the data set to double the number of cases. They plan to conduct a prospective study that investigates how radiologists interact with the model's prediction, such as how they use it for interpretation.

When asked if this model has the potential to detect other types of cancer, Takahashi explained that since it's specifically trained using prostate MR, it's unlikely to detect cancer outside the prostate gland. However, if they were to train a different model using the same approach, it would likely be able to detect other types of cancer.

Artificial Intelligence Homepage