Click here to

Session: AI in Imaging [Return to Session]

Explainable GRU-Survnet Using Longitudinal Quantitative Imaging Biomarkers From MRI and PET for Local Failure in Poor-Prognosis Head and Neck Cancer

L Wei*, M Mierzwa, C Lee, Y Cao, The University of Michigan, Ann Arbor, MI


TH-D-207-4 (Thursday, 7/14/2022) 11:00 AM - 12:00 PM [Eastern Time (GMT-4)]

Room 207

Purpose: To investigate a gated recurrent unit-based survival neural network (GRU-Survnet) for prediction of local failure (LF) of poor-prognosis head and neck cancers (HNCs) based upon longitudinal multi-modality quantitative imaging biomarkers (QIBs); and understand underlying feature attributions and interactions.

Methods: 93 patients with p16+ T4/N3 squamous cell carcinoma of oropharynx or locally advanced p16– HNCs enrolled in a randomized phase II adaptive radiation dose escalation trial were studied. FDG-PET and MRI scans were acquired pre-RT and at fraction 10 of RT. MRI-derived gross tumor volume (GTV), blood volume (BV) and apparent diffusion coefficients (ADC), as well as FDG metabolic tumor volume (MTV50) features were analyzed in primary tumors. A GRU-Survnet that could incorporate correlative signals from two time points was investigated for prediction of LF using p16 status, boost and QIBs. Individual feature attributions and feature interactions in the GRU-Survnet were evaluated by integrated gradients (IG) and Expected Hessian (EH), respectively. Cox model as well as a simple fully connected network (Deepsurv) with direct inputs of all features were used as benchmarks. 10 times 5-fold cross-validation was used to avoid overfitting. Concordance indexes (c-index) were compared between models using ANOVA.

Results: The GRU-Survnet achieved c-index of 0.758 (0.726-0.790) compared with 0.671 (0.607-0.734) of Cox model and 0.723 (0.686-0.760) of Deepsurv. GRU-Survnet was significantly better than the Cox model (p<0.05). P16 status, boost, GTV, subvolumes of low BV and low ADC volume are top-ranked biomarkers. EH analysis showed that p16 status had the most interaction with QIBs (top-ranked features being ADC-based) and boost is more important for p16+ patients.

Conclusion: An interpretable GRU-Survnet enabled better prediction of LF in poor-prognosis HNCs using MRI and FDG-PET biomarkers. Biomarker attribution and interaction studies revealed interesting interplay patterns of p16 status, boost and QIBs. These findings could aid in for personalized local intensified treatment.

Funding Support, Disclosures, and Conflict of Interest: NIH R01 CA184153 NIH U01 CA183848


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email