Deep learning in head & neck cancer outcome prediction

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Traditional radiomics extracts quantitative texture features from medical images to correlate with clinical outcomes.
• Limitations exist in traditional radiomics for capturing all relevant image patterns.

Purpose of the Study:

• To investigate if convolutional neural networks (CNNs) can enhance traditional radiomics performance.
• To predict treatment outcomes in head and neck squamous cell carcinoma using pre-treatment CT images.

Main Methods:

• Trained a CNN on pre-treatment computed tomography (CT) images from 194 patients with head and neck squamous cell carcinoma.
• Validated the CNN on an independent set of 106 patients from 4 institutions.
• Compared CNN performance against a traditional radiomic framework.

Main Results:

• The CNN achieved an Area Under the Curve (AUC) of 0.88 in predicting distant metastasis.
• Combining the CNN model with the traditional radiomic model improved the AUC to 0.92.
• The CNN models recognized traditional radiomic features and were visualized.

Conclusions:

• CNNs can enhance traditional radiomics by identifying image patterns beyond conventional feature extraction.
• The combined approach offers improved accuracy for predicting distant metastasis in head and neck cancer.
• This framework provides interpretable and accurate outcome prediction models.