An AI neural network can accurately predict the prognosis of melanoma patients based on pre-treatment histology imaging data, shows research led by the NYU Grossman School of Medicine.
Immune checkpoint inhibitors have revolutionized melanoma treatment, but only some tumors respond well to them and they can be quite toxic to patients.
Having a more reliable way to predict who is most likely to respond to these therapies is therefore crucial.
“An unmet need is the ability to accurately predict which tumors will respond to which therapy,” says Iman Osman, M.D., a medical oncologist based at New York University (NYU) Grossman School of Medicine and NYU Langone’s Perlmutter Cancer Center, who co-led the work.
“This would enable personalized treatment strategies that maximize the potential for clinical benefit and minimize exposure to unnecessary toxicity.”
In collaboration with Aristotelis Tsirigos, Ph.D., professor in the Institute for Computational Medicine at NYU Grossman School of Medicine and member of NYU Langone’s Perlmutter Cancer Center, Osman and team first trained an artificial neural network using pre-treatment histology images from 121 patients with metastatic melanoma.
As reported in the journal Clinical Cancer Research, all the patients were treated with immune checkpoint inhibitors (anti-CTLA-4, anti-PD-1, or a combination) between 2004 and 2018 and their responses were recorded.
After the neural network was trained, the team tested the accuracy of its predictions regarding prognosis using images from 30 additional patients with metastatic melanoma. These individuals were treated with the same immune checkpoint inhibitor drugs at Vanderbilt-Ingram Cancer Center between 2010 and 2017.
The researchers used a statistical test called area under the curve, or AUC – where a score of 0.7-0.8 is satisfactory, 0.8-0.9 is excellent and over 0.9 is exceptional — to assess how well the network could predict treatment outcome.
They found that the network scored approximately 0.7 in both the training and validation groups of patients. However, when other clinical features such as age, gender, and histologic subtype, among others, were added into the model the predictive AUC score improved to 0.8.
“Several recent attempts to predict immunotherapy responses do so with robust accuracy, but use technologies, such as RNA sequencing, that are not readily generalizable to the clinical setting,” said Tsirigos.
“Our approach shows that responses can be predicted using standard-of-care clinical information such as pre-treatment histology images and other clinical variables.”
The researchers say that a next step for them is to train the neural network using a much larger set of images, as less than 400 were used in this study.
“There is data to suggest that thousands of images might be needed to train models that achieve clinical-grade performance,” said Tsirigos.