Generating prior probabilities for classifiers of brain tumours using belief networks

Area of Science:

• Neuro-oncology
• Medical imaging
• Machine learning

Background:

• Supervised machine learning is commonly used for brain tumor classification based on MRI spectra and imaging.
• Few studies have explored multimodal classification combining imaging and spectroscopy.
• This work presents a novel method for generating tumor class probabilities from anatomical location.

Purpose of the Study:

• To introduce Bayesian belief networks for generating tumor type probabilities.
• To demonstrate the application of belief networks using a pediatric tumor dataset.
• To evaluate the improvement in classification accuracy by combining anatomical and MRS data.

Main Methods:

• Bayesian belief networks were constructed using a five-decade pediatric tumor database (>1300 cases).
• The method generates probabilities for various tumor types, including specific astrocytoma grades, ependymoma, pineoblastoma, PNET, germinoma, medulloblastoma, craniopharyngioma, and rare tumors.
• Prior probabilities from anatomical location were generated and combined with Magnetic Resonance Spectroscopy (MRS) data for classification.

Main Results:

• Belief networks successfully generated probabilities for classifying diverse brain tumor types.
• Using the network to generate prior probabilities enhanced classification accuracy compared to using class prevalence alone.
• The developed networks provide probabilities for astrocytoma (grades I-IV), ependymoma, pineoblastoma, PNET, germinoma, medulloblastoma, craniopharyngioma, and 'other' rare tumors.

Conclusions:

• Bayesian belief networks offer a straightforward method for integrating discrete clinical data to generate classification probabilities.
• The belief network approach demonstrates robustness with incomplete datasets.
• Incorporating prior knowledge through belief networks effectively improves non-invasive brain tumor classification.