Developing a robust two-step machine learning multiclassification pipeline to predict primary site in head and neck...

Area of Science:

• Radiology and Oncology
• Medical Imaging Analysis
• Machine Learning in Medicine

Background:

• Head and neck cancers of unknown primary present diagnostic challenges.
• Accurate primary tumor site identification is crucial for effective treatment strategies.

Purpose of the Study:

• To develop and evaluate a robust multiclassification pipeline using radiomics and machine learning to determine primary tumor locations in head and neck cancers of unknown primary.
• To compare a direct identification pipeline (P1) with a two-step approach (P2) for improved accuracy.

Main Methods:

• A dataset of 400 head and neck cancer patients (oropharynx, nasopharynx, oral cavity, larynx/hypopharynx) was analyzed.
• Two radiomic-based multiclassification pipelines (P1 and P2) were developed and compared.
• Diverse feature selection methods and classification models, including support vector machines, were assessed.

Main Results:

• The two-step pipeline (P2) outperformed the direct identification pipeline (P1).
• The best performance was achieved using a support vector machine classifier with radiomic and clinical features.
• High accuracies were reported for specific sites: 75.3% (larynx/hypopharynx), 75.4% (oral cavity), 71.3% (oropharynx), and 92.9% (nasopharynx).

Conclusions:

• The two-step multiclassification pipeline integrating radiomics and clinical information is a promising approach for predicting primary tumor sites in unknown primary head and neck cancers.
• This method has the potential to aid in diagnosis and guide treatment decisions for these challenging cases.