A chest CT taken before immunotherapy can already hint at which lung cancer patients will suffer a serious inflammatory reaction. Researchers at MD Anderson trained a deep learning model to read those clues.
The condition, pneumonitis, strikes roughly one in ten lung cancer patients treated with immune checkpoint inhibitors. It is hard to predict and can escalate into a life-threatening emergency.
The model is called CIPHER. It was tested on pretreatment scans from 347 people with non-small cell lung cancer at MD Anderson, then run against a separate external dataset. Both cohorts produced an area under the curve near 0.83. Models built from clinical risk factors and hand-designed radiomics features scored lower.
That consistency mattered. The validation group differed in patients, scanner types and imaging protocols, the kinds of shifts that usually degrade image models. CIPHER held its accuracy anyway.
High-risk patients also developed pneumonitis sooner after starting treatment. That timing suggests the model tracks a continuous gradient of lung vulnerability rather than sorting people into two buckets.
Jia Wu, an associate professor of imaging physics who led the work, said the model was never told to hunt for pneumonitis. Instead it picked up subtle lung tissue patterns linked to later toxicity, which suggests routine scans carry more information about treatment harm than clinicians had assumed.
The study appears in the Journal for ImmunoTherapy of Cancer. Prospective studies in larger, more diverse groups come next, along with tests in other cancers treated with immunotherapy.
