Deep Learning Model to Predict Dry Mouth in Head and Neck Cancer Patients
Author Information
Author(s): Chu Hung MSc, de Vette Suzanne P.M. MSc, Neh Hendrike MSc, Sijtsema Nanna M. PhD, Steenbakkers Roel J.H.M. MD, PhD, Moreno Amy MD, Langendijk Johannes A. MD, PhD, van Ooijen Peter M.A. PhD, Fuller Clifton D. MD, PhD, van Dijk Lisanne V. PhD
Primary Institution: University Medical Center Groningen, University of Groningen, Groningen, The Netherlands
Hypothesis
Using 3D dose distribution, CT imaging, and clinical variables can improve the prediction of late xerostomia risk in head and neck cancer patients.
Conclusion
Deep learning models for predicting xerostomia performed better than traditional models when validated on internal data, but struggled with external validation.
Supporting Evidence
- Deep learning models showed better performance than traditional models in predicting xerostomia.
- Attention maps indicated that the model focused on the major salivary glands.
- Transfer learning improved model performance on external validation data.
- Models were trained on a large cohort of 1208 patients.
- Statistical evaluation metrics included AUC, Brier score, and Nagelkerke R2.
Takeaway
This study used advanced computer models to better predict dry mouth in cancer patients after treatment, showing that using more detailed information helps make better predictions.
Methodology
The study trained deep learning models using 3D dose distributions, CT scans, and clinical variables from a cohort of 1208 patients.
Potential Biases
Differences in scoring systems and data acquisition methods between centers may introduce bias.
Limitations
The models did not generalize well to external validation cohorts, indicating a need for multicenter training data.
Participant Demographics
Patients included were diagnosed with head and neck cancer, with a mix of genders and ages.
Statistical Information
P-Value
0.04
Confidence Interval
0.72–0.78
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website