Last year, we were fortunate enough to embark on a groundbreaking AI research project in collaboration with Stanford University to leverage our NeoPulse platform to apply artificial intelligence (AI) to brain PET/CT scans with stellar results.
As we discussed in our blog detailing the AI research, this type of technology advancement will be monumental for oncology and diagnostic imaging. Bringing AI into clinics will assist medical professionals to help reduce the burden on them and allow for exponential scale resulting in rapid and greater medical care to more people.
And this is precisely why we’ve continued to collaborate with Stanford to determine how deep learning (DL) and AI can advance the field of oncology with the hopes of bringing the technology to life within clinics around the world in the very near future.
In fact, we’ve just published another research paper* detailing these continued advancements – this time to apply deep learning to detect prostate cancer recurrence. Using similar methodology as the previous research (read full paper for precise details), the researchers applied deep learning via the NeoPulse platform to PET studies to classify them as normal or abnormal.
However, what makes this study different than the past one we did, and indeed different than previous studies in the field, is the method that was used. The researchers compared three different approaches – 2D-CNN slice-based approach, 2D-CNN case-based approach and 3D-CNN case-base approach. According to the paper, it handled images with “a large field of view, including numerous transaxial PET slices and a wider spectrum of abnormality and normality than that used in previous studies. The slices used in this study had variations including the number, shape, size, and location of local recurrence of PC, as well as uptake in lymph nodes or distant metastases.”
Similar to the previous study though, PET scans were labeled as either normal, abnormal, or indeterminate where the normal and abnormal scans were utilized as training and test datasets. For this particular study, there were 170 abnormal, 64 normal and 17 indeterminate PET images. The research team selected the training dataset to consist of 100 PET images – 50 normal and 50 abnormal – and a test dataset of 28 PET images – 14 normal and 14 abnormal – that were randomly distributed. Just like the last study and just as a radiologist would do, we trained the models for a variety of different ways to view the scans and make a classification.
And we couldn’t be more excited by the results…
Out of the 3 tested approaches, the study concluded that NeoPulse accurately classified 2D-CNN images at a rate of 89.24% accuracy. In fact, the researchers involved in this project were left with this key takeaway that we believe is extremely promising for the future of medicine:
“Thinking about a potential deployment of models like this at the point-of-care is of importance to the authors. A variable of value in such clinical setting is time. The processing of 28 test PET images and 536 test transaxial slices took less than 1 min. This suggests that once a model is properly trained and available in a clinical workflow, a probability ‘score of abnormality’ could be provided in real time following PET scan reconstruction. A probability ‘score of abnormality’ alone may help prioritize reads and/or increase reader confidence.”
This latest research has us incredibly optimistic that we’ll start seeing artificial intelligence in partnership with radiologists and nuclear medicine physicians in clinics to help change the way healthcare professionals screen for cancer, resulting in faster time to diagnosis and paving the way for greater scalability.
This research demonstrates powerful potential that with the assistance of AI, radiologists may one day be able to deliver real-time results at the point-of-care.
*You can read the final research paper here.
Jong Jin Lee, & Hongye Yang, & Benjamin L. Franc, & Andrei Iagaru, & Guido A. Davidzon. (2020). Deep learning detection of prostate cancer recurrence with 18F-FACBC (fluciclovine, Axumin®) positron emission tomography. European Journal of Nuclear Medicine and Molecular Imaging. https://doi.org/10.1007/s00259-020-04912-w
