American Association for Cancer Research
crc-22-0152_fig3.png (157.28 kB)

FIGURE 3 from Multi-institutional Prognostic Modeling in Head and Neck Cancer: Evaluating Impact and Generalizability of Deep Learning and Radiomics

Download (157.28 kB)
posted on 2023-06-29, 14:20 authored by Michal Kazmierski, Mattea Welch, Sejin Kim, Chris McIntosh, Katrina Rey-McIntyre, Shao Hui Huang, Tirth Patel, Tony Tadic, Michael Milosevic, Fei-Fei Liu, Adam Ryczkowski, Joanna Kazmierska, Zezhong Ye, Deborah Plana, Hugo J.W.L. Aerts, Benjamin H. Kann, Scott V. Bratman, Andrew J. Hope, Benjamin Haibe-Kains

Volume dependence of predictions. Spearman rank correlation of the predictions of each model with tumor volume against performance in terms of AUROC (A) and C-index (B), respectively. The top models fall into an optimal region of low (but nonzero) volume correlation and high performance. Note that while models 1 and 2 used tumor volume as one of the input variables, their predictions correlate with volume only moderately (ρ < 0.5), indicating they are able to exploit additional information present in the EMR features. Higher correlation leads to decreased performance as the predictions are increasingly driven by volume only. Most radiomics-only models fall in the high correlation region (ρ ≥ 0.5), although deep learning predictions correlate at notably lower level than engineered features. Interestingly, the best radiomics submission (number 9) achieves the lowest volume correlation, suggesting that it might be using volume-independent imaging characteristics.


Canadian HIV Trials Network, Canadian Institutes of Health Research (CTN, CIHR)



ML combined with simple prognostic factors outperformed multiple advanced CT radiomics and deep learning methods. ML models provided diverse solutions for prognosis of patients with HNC but their prognostic value is affected by differences in patient populations and require extensive validation.