A new AI tool can analyze whole-body PET/CT scans and detect six different types of cancer, according to research presented on Monday at the SNMMI Annual Meeting in Toronto. In the future, this tool can help physicians assess a patient's risk, predict their treatment response, and estimate their survival.

"There is a need for an AI tool like the one developed in the present study as there are, to my knowledge, no existing tools that automatically perform both tumor burden quantification and prognostic prediction on PET/CT scans of patients with cancer," Kevin H. Leung, research associate at Johns Hopkins University School of Medicine, told HCB News.

He added that existing software solutions may offer semi-automated approaches for tumor delineation or manual regions of interest definition, but they are often limited in their performance and generalizability and may require many user-defined inputs to produce acceptable results.

Leung and his team created the tool using a type of AI called deep transfer learning, which uses the knowledge gained from one task to improve the performance of another task. The result is a tool that can fully automate, whole-body tumor segmentation and prognosis on PET/CT scans.

To test its effectiveness, the team conducted a study using data from 611 FDG PET/CT scans of patients with melanoma, lymphoma, lung, head and neck, and breast cancer. They also included 408 PSMA PET/CT scans of prostate cancer patients.

The AI tool automatically extracted radiomic features and whole-body imaging measures and was able to quantify molecular tumor burden and uptake from the predicted segmentations. That information was used to develop prognostic models that stratify the risk of prostate cancer, estimate the survival of head and neck cancer patients, and predict the treatment response of breast cancer patients undergoing neoadjuvant chemotherapy.

According to Leung, not only can this tool determine cancer prognosis, but it also holds the potential to improve patient outcomes and survival because it can identify robust predictive biomarkers, characterize tumor subtypes, and ultimately help with the early detection and treatment of cancer.

It might also be able to provide value in the early management of patients with advanced, end-stage disease because it can identify the right treatment and predict their responses to therapies like radiopharmaceutical therapy.

"The developed AI tool, and other similar tools, will be integrated into the radiologists’ workflow seamlessly," explained Leung. "As soon as an image is loaded onto a workstation, the developed approach will automatically perform tumor segmentation and extract radiomic features and whole-body imaging measures from the predicted segmentations to quantify molecular tumor burden and uptake."

Other AI tools that detect cancer are built using small- to moderately-sized data sets that typically involve a single malignancy or radiotracer. Leung referred to this new model as a "critical bottleneck in the current training and evaluation paradigm for AI and machine learning applications in medical imaging and radiology."

He believes that fully-automated AI tools like this will one day be a staple in imaging centers, helping physicians interpret cancer patients' PET/CT scans. They will be able to automatically detect lesions, delineate volumes of interest, and quantify important features and tumor characteristics that are difficult to evaluate visually.

"This automatic quantitative analysis will reduce the overall physician workload and augment radiologist capabilities by acting as a secondary reader for whole-body imaging scans," said Leung.

In addition to the clinical applications, this tool may also better the understanding of different diseases and cancers. That's because it can provide important molecular insights about the underlying in vivo biological processes and phenomena that might currently be understudied in large-scale populations.

Leung said that plans for making this tool widely available are currently underway.