Using AI to screen breast cancer

Policymakers, publics, and practitioners have great aspirations for the potential of AI to improve healthcare systems and patient outcomes. Delivering on these aspirations requires trustworthy AI that can be integrated into clinical practice, leveraging methods for the analysis of multimodal data sources and explainable AI to create effective decision-support systems.

Breast cancer is the most common type of cancer among women, with one in eight diagnosed during their lifetime.  When the disease is detected early, for example through routine mammograms, treatment can be highly effective, achieving survival probabilities of 90% or higher.  However, mass screening increases radiologists’ workloads, generating more images to analyse. Despite widespread digitization of images and clinical records, collecting data to build time-saving tools is not easy: the quality of images from screenings varies, and there are complex privacy rules to navigate.

Intelligent diagnoses
AIGEA Medical is applying AI to digital imaging workflows in radiology with the aim of helping radiologists diagnose breast cancer more quickly by speeding up repetitive tasks.  “If we can use AI to screen out negative cases and bring suspicious or positive cases to the attention of radiologists, it could reduce the time it takes to get a diagnosis from weeks to hours,” says Carlo Aliprandi, cofounder and CEO of the Italian MedTech start-up.

The company says its AI cloud-based software, called DeepMammo, increases the accuracy of diagnoses because it works from multiple sources of data. Its patented ‘Multimodal Learning AI’ based on Deep Neural Networks is applied to a dataset of exams, combining images from screening with text from clinical reports. “We’re not just using images to train DeepMammo, but data from clinical records, such as age and tissue density,” Aliprandi explains.

At first, the start-up focused on processing medical images, using AI algorithms to analyse scans and generate a classification of positive or negative for breast cancer by exploiting state-of-the-art Convolutional Networks (CNNs). Connecting these capabilities to clinical practice required additional functions, specifically to take into account other clinical inputs and to present outputs in a way that supported clinicians in their work. Its work on the ELISE programme allowed AIGEA Medical to work on explainability, using Natural Language Generation (NLG) to describe important features of the images and incorporate multimodal data, increasing the power and flexibility of DeepMammo.

The company intends for its technology to plug into clinical systems already used by radiologists to manage their workflow. The idea is to use AI to screen a patient’s scans and records before or at the point when a radiologist is due to make a reading of the examination.

ELISE helped AIGEA Medical implement a core step in its roadmap, enriching DeepMammo with a novel AI for automatic generation of medical reports from images, using NLG technology. “With the support of ELISE and the company’s mentor, Professor Eneko Agirre of the Computer Science Faculty of the University of the Basque Country UPV/EHU, we could move along with this step. It was very useful for us,” Aliprandi says. The novel AI can provide the radiologist with a positive or negative classification for the presence of the signs of cancer, but also a useful report that includes data to inform clinical decision-making, such as relevant details in notes or images descriptions that they can use. DeepMammo now explains how it arrives at a diagnosis and provides transparency to the user. “That’s what we mean by explainability – explaining a diagnosis to radiologists in terms that are meaningful to them,” Aliprandi says.

Visualising the future
“We are at the very beginning of the age of AI applied to digital imaging diagnostics,” Aliprandi predicts. “Work has to be done in order to fully face the ethical and regulatory challenges of AI in healthcare, particularly privacy, data fairness and trustworthiness. Addressing these issues is integral for a solution like DeepMammo, which aims to leverage technology in the interest of democracy and common well-being.” He hopes his company’s technology and other AI in the wider community will be applied to diagnose different diseases and be used for everything from mammograms to MRI scans. “There are technical challenges, including adapting AI techniques for different modalities and diseases, as well as collecting the necessary amount of data.” Aliprandi believes that eventually, AI will be a big support for radiologists and other clinicians, allowing them to leave mundane work to algorithms and focus on tasks that really add value, such as diagnosing complex cancer types or delivering better patient care.