Artificial intelligence can help in the non-invasive diagnosis of prostate cancer

An international and multidisciplinary team, led by researchers from the Champalimaud Foundation, in Lisbon, has developed an artificial intelligence algorithm that improves the reliability of prostate cancer detection based solely on magnetic resonance imaging (MRI) scans – that is, without resorting to a biopsy. Their findings were published today, 15th of August, in the journal Radiology: Imaging Cancer.

The standard diagnostic protocol for prostate cancer begins when a man has a routine PSA (prostate-specific antigen) blood test and the value of this indicator is abnormally high. The man then undergoes an MRI scan to determine whether or not there are any suspicious lesions in his prostate.

Radiologists then analyse the potentially cancerous lesions in the images obtained and summarise their assessment using a scale of 1 to 5, called PI-RADS (Prostate Imaging Reporting and Data System), which takes into account various visual characteristics of the lesions, such as texture, shape, and size. PI-RADS values of 1 and 2 represent a low risk of cancer, so active surveillance of the lesions is generally recommended, without additional interventions, to monitor any changes that may indicate progression to malignancy. Values of 4 or 5 mean that the presence of cancer is very likely. A biopsy is generally recommended to confirm the diagnosis, and if it confirms that the lesions are indeed cancerous, the most appropriate treatment – surgery, radiation therapy, chemotherapy, immunotherapy, etc. – will be chosen on a case-by-case basis.

If the PI-RADS score is 3, the need for a biopsy to rule out cancer depends on factors such as the location of the lesion, PSA levels, and other clinical data.
However, when the PI-RADS is greater than 3, a biopsy is almost always performed, since even when the imaging results indicate the absence of cancer, this may turn out to be a “false negative.”

Biopsies are invasive procedures, which may involve multiple prostate punctures to remove cell samples from the lesions. This can cause a lot of discomfort for patients, as well as infections – not to mention the additional costs that the test entails.

José Almeida (a postdoctoral researcher specialised in artificial intelligence and data science) and the FC's Computational Clinical Imaging Group, directed by Nickolas Papanikolaou, lead author of the study – in collaboration with colleagues from other institutions in Portugal, Greece, and Italy – wanted to know if AI could help reduce the number of unnecessary biopsies. Would it be possible to reliably predict whether lesions visible on MRI images are malignant or not? “Our model attempts to predict whether the biopsy will be positive or not, directly from the radiological diagnosis,” explains Almeida.

The study now published concludes that the use of AI in the radiological assessment of prostate lesions can help radiologists make a robust diagnosis without resorting to biopsies in about 20 percent of cases.

A wealth of data

The work was based on magnetic resonance imaging data of the prostate collected over several years as part of the ProCancer-I project. Launched in 2020 by a European consortium, it was funded by the European Union's Horizon 2020 program (https://www.fchampalimaud.org/pt-pt/news/inteligencia-artificial-pode-e…).

The project focused on improving radiological diagnosis and also on verifying whether there was room for further improvement by using artificial intelligence models to perform radiological diagnoses of prostate cancer.

As part of ProCancer-I, a centralised image archive, ProstateNET, was developed, containing tens of thousands of prostate MRI scans. Coming from 12 clinical centres, nine countries, and three MRI machine manufacturers, this data was intended to be used to train and test AI models. “Thanks to this platform, we were able to have a volume of cases that is difficult to match – close to 9,000 – to train our model,” says Almeida.

The information provided to the model during the training phase included MRI images, the patient's age, PSA value, PI-RADS value, and the location of the lesion in the prostate. “All of these variables were available at the time the patient underwent the radiological examination,” Almeida points out. “We did not give the model anything that it would not have access to in a real situation.”

Some of the data, classified as “retrospective” – namely, collected before March 31, 2022 – was used to train the AI model. Another part – the “prospective” data, recorded after that date – was used to validate the model. This temporal separation allowed the robustness of the model's predictions to be tested.
In addition, the AI model was also tested with a set of data from Agios Savas Hospital (Greece) that was not included in the image archive, in an “external prospective validation with data completely unknown to the model,” as Almeida explains.

Finally, it should be noted that the model was tested on images obtained with multiple MRI modalities and, as already mentioned, different brands of machines.

Fewer unnecessary biopsies

The main conclusion of the study is that the AI model demonstrated greater sensitivity (ability to identify biopsies that are actually positive) and specificity (ability to identify biopsies that are actually negative) than the PI-RADS scale.

“The model outperformed PI-RADS in retrospective validation (...) and in prospective validation sets, leading to 22.7% fewer biopsies compared to PI-RADS”, the authors write in their paper.

On the other hand, “sensitivity analyses showed the importance of multiple sequences [several simultaneous MRI imaging techniques] and multiple [machine] suppliers (...), since the use of specific sequences or suppliers alone led to poorer performance.”

Part of the study also involved verifying the model's performance in various scenarios, as the limitations of these models must be detected as early as possible to prevent them from being used in circumstances that could lead to diagnostic errors. For example, when the area of the body covered by the exam is too large, the model is more likely to fail. Knowing this makes it possible to apply the model more systematically and safely.

The application of this model to clinical practice may take a few years. “The developed model,” the authors write, “has the potential to help physicians reduce the number of unnecessary biopsies, as it demonstrated extremely high generalisability and proved robust in most of the analysed sub-cohorts. Future work should focus on incorporating additional sources of clinical information, such as race/ethnicity, family history of prostate cancer or other hereditary cancer syndromes, and genetic risk for prostate cancer.” It will also be necessary to better understand the model’s performance in real-world settings.

“The idea is to continue validating this approach in a more robust and systematic way,” says Almeida. “Big claims require big evidence.”
 

Original paper here.

Text by Ana Gerschenfeld, Health & Science Writer of the Champalimaud Foundation.