Deep Learning-Based Estimation of Tumor Microstructure Parameters in Head and Neck Cancers

Motivation: Microstructure parameter mapping using the random walk with barrier model shows promise for head and neck cancer (HNC) diagnosis and prognosis but lacks robustness in clinical settings. Goal(s): To improve accuracy and precision of parameter fitting by deep learning (DL). Approach: We evaluated and compared conventional and DL methods on simulated data (estimation squared error, bias, variance) and in vivo HNC data (histogram, visual analysis). Results: In simulated data, DL methods reduced estimation squared error and variance across different noise levels and reduced bias at high noise levels. In vivo, DL methods decreased noise and mitigated unrealistic estimates. Impact: The improvement in parameter quantification accuracy and precision of the random walk with barrier model by DL methods will facilitate a clinical application in anti-cancer treatment response assessment in head and neck cancers, potentially benefiting adaptive therapy.