Detection of extraprostatic extension of prostate cancer by transfer learning AI models versus radiologists on biparametric and multiparametric MRI

Prostate cancer is the most common cancer among men in the United States, and determining how far the disease has spread is a critical step in deciding how to treat it. When a tumor is confined within the prostate gland, doctors can often remove it with surgery that preserves important nerves. However, if the cancer has pushed past the gland's outer boundary, a condition known as extraprostatic extension, the treatment plan must change. Surgeons may need to remove wider margins of tissue, and patients might require additional therapies like radiation or hormone treatment to prevent the cancer from returning. For decades, doctors have relied on magnetic resonance imaging, or MRI, to spot these signs of spread before surgery. While these scans are powerful, reading them is difficult. The signs of a tumor breaking through the gland's edge can be subtle, and even experienced specialists sometimes miss them or disagree on what they see.

To address this challenge, a team of researchers at NYU Grossman School of Medicine and the University of Washington set out to see if artificial intelligence could match the skill of human radiologists in detecting this specific type of cancer spread. They focused on two types of MRI scans: a standard version that uses multiple types of data, and a streamlined version that uses fewer sequences but is faster and cheaper to perform. The researchers built computer models capable of learning from thousands of past patient scans to recognize the patterns associated with cancer that has extended beyond the prostate. Their goal was not just to see if the machines could learn, but to determine if they could do so without needing a human to first draw detailed outlines around the prostate or the tumor, a process that is slow and prone to human error.

The study involved a large group of patients who had undergone prostate MRI scans between 2015 and 2023 and subsequently had their prostate removed surgically. This surgical removal provided the ground truth, allowing the researchers to know exactly which patients had cancer that had spread and which did not. The team divided this data into groups to train and test their computer models. They created two versions of an artificial intelligence system. The first version looked only at the images from the streamlined MRI, which captures the structure of the prostate and how water molecules move through it. The second version looked at those same images plus a third type of scan that shows how the prostate takes up contrast dye. Crucially, neither version required a human to manually outline the prostate or the tumor on the screen. Instead, the computer analyzed the entire volume of the scan at once, learning to identify the subtle textures and shapes that indicate cancer has breached the gland's boundary.

To see how well these computer models performed, the researchers compared them against three highly trained radiologists who specialize in abdominal imaging. These experts reviewed the same test scans, first looking at the streamlined images and then at the full set of images including the contrast dye. They assigned a score to each scan based on how likely it was that the cancer had spread. The results showed that the computer models performed just as well as the human experts. The models were able to detect the spread of cancer with a level of accuracy that was statistically indistinguishable from the fellowship-trained radiologists. This finding is significant because it suggests that artificial intelligence can serve as a reliable second pair of eyes, potentially reducing the workload for doctors and ensuring that fewer cases of spread are missed.

The researchers also investigated whether adding the contrast dye images helped the diagnosis. For both the human readers and the computer models, looking at the full set of images including the dye led to a slight improvement in accuracy. The humans became better at spotting the spread, and the computer models became slightly more precise. However, the improvement was not dramatic enough to declare it a definitive game-changer based on this data alone. The study also found that the computer models were particularly good at spotting larger areas of spread, while both the humans and the machines struggled more with very small, focal areas where the cancer had just begun to poke through the boundary. This difficulty with small, focal cases is a known limitation in the field, as these tiny signs are inherently harder to see on any scan.

One of the most notable aspects of this work is how the computer models were trained. Unlike many previous attempts that required doctors to spend hours drawing precise borders around the prostate and the tumor before the computer could learn, these models learned directly from the raw images and the final surgical results. This approach, known as weakly supervised learning, means the system learned to recognize the disease without the bottleneck of manual outlining. The researchers also tested whether adding a simple blood test measurement, known as prostate-specific antigen density, to the computer's analysis would help. While adding this clinical data did improve the model's performance slightly, the images themselves remained the primary driver of the diagnosis.

The study concludes that these segmentation-free artificial intelligence models are a viable tool for detecting cancer spread, performing on par with expert radiologists. While the addition of contrast dye and clinical data offered small gains, the core finding is that the machines can learn to see what the experts see without needing the time-consuming step of manual tracing. The researchers acknowledge that their work was done at a single medical center and that the models need to be tested on data from different hospitals and different types of MRI machines to ensure they work everywhere. They also noted that the study was retrospective, meaning it looked back at past data, and that future work should focus on refining the models to catch the smallest, most difficult cases of spread. For now, the work demonstrates that artificial intelligence has reached a point where it can stand shoulder-to-shoulder with human specialists in a high-stakes diagnostic task, offering a promising path toward more consistent and accessible cancer care.