Screening for extranodal extension using AI
In an article in the Lancet Digital Health, we demonstrate the value of a deep learning algorithm able to screen for extranodal extension (ENE)in head and neck cancers. This validation study was conducted using data from ECOG-ACRIN Cancer Research Group E3311, a multinational trial wherein HPV and OPC were treated surgically and assigned to a pathologic risk-based adjuvant strategy of observation, radiation, or concurrent chemoradiation.
We evaluated the deep learning algorithm to predict ENE and found that it performed well, with an area under the receiver operating characteristic curve (AUC) of 0.85. This was comparable to performance on prior validation sets (0.84 – 0.90) and outperformed four head and neck radiology experts (AUC range: 0.63 – 0.71), with diagnostic gains largely due to increase sensitivity. Coupled with rigorous resiliency testing and uncertainty estimation, the study demonstrates that deep learning could improve diagnosis of ENE compared to the standard of care. We have released the model and source code to facilitate independent validation of the model and promote prospective testing.
AIM Investigators
