Prognostic modeling in head and neck cancer: deep learning or handcrafted radiomics?

Abstract Head and neck squamous cell carcinoma (HNSCC) presents a complex clinical challenge due to its heterogeneous nature and diverse treatment responses. This review critically appraises the performance of handcrafted radiomics (HC) and deep learning (DL) models in prognosticating outcomes in HNSCC patients treated with (chemo)-radiotherapy. The focus was on methodological rigor, performance metrics, and long-term outcome reporting. A comprehensive literature search was conducted up to May 2023, identifying 23 eligible studies that met the inclusion criteria. We analyzed methodological variability and predictive performance metrics of both models for outcomes including overall survival, loco-regional recurrence, and distant metastasi. Our findings concluded that DL models exhibited slightly superior performance metrics compared to HC models, particularly in outcome prediction. However, the highest methodological quality was noted predominantly in HC studies. Substantial variability in methodology, outcome definitions, and performance metrics was observed, highlighting the need for standardization. While DL models show potential for improved prognostic performance, the methodological robustness in HC studies underscores their reliability. This emphasizes the necessity for methodological improvements, including pre-registration of protocols and clinical utility assessments, to enhance the reliability and applicability of radiomics-based prognostic models in clinical practice.