Decoding Brain Fog in Head and Neck Cancer Survivors Using Artificial Intelligence

Purpose/Objective(s)

Brain fogging, which is characterized by memory decline, word-finding difficulty, and decreased multi-tasking ability, is common in head and neck cancer survivors. The precise causes of brain fog are poorly understood, as MRIs are essentially normal in these patients. The purpose of this study was to use artificial intelligence (AI) to detect micro-architectural and biological changes in the brain that are invisible to human eyes.

Materials/Methods

DVH of temporal lobes and whole brain were obtained in 566 and 60 patients, respectively, who underwent IMRT or proton radiation with concurrent chemotherapy for nasopharyngeal carcinoma. EORTC QLQ-C30 was obtained from 92 patients who enrolled in two prospective clinical studies. Automated cortical region segmentation using an open-source software for processing and analyzing brain MRI images was performed in 33 patients with a minimum follow-up of 4 years. Architectural and biological MRI changes in gray matter, white matter, and subcortical brain regions were determined in 34 brain regions. Graph theory-based brain connectome was evaluated to assess the connectivity between different regions of the brain. Radiomics which employs high-throughput quantitative analysis of image features were performed on the T1- and diffusion-weighted sequences.

Results

Patients experienced a continuous decline of cognitive function after radiation, with a 6.6% in decrease in their first year. The brain received a large low-dose bath with a median V10-20(whole brain) (volume of whole brain that received 10-20 Gy) 170cc, V20-30(whole brain) 50.8cc, V10-20(temporal lobe)11.4cc, and V20-30(temporal lobe) 6.18cc. There were global changes in gray matter thickness, with most pronounced changes occurred in parietal lobe (-4.79%, p=0.007) and occipital lobe (-5.68%, p=0.03). Similarly, there were diffuse changes in white matter and subcortical volume. After radiation, the frontal lobe increased by 17.5% (p=0.04), lateral ventricle 41% (p=0.03), and choroid plexus 34.3% (p=0.03). After radiation, the brain exhibited diffuse disrupted connectivity as characterized by increased clustering coefficient (p=0.002) and decreased modularity (p=0.003). Corpus collusum, choroid plexus, cerebellum, hippocampus, brainstem, medialorbito- and superior-frontal, temporal pole, and inferior temporal were analyzed using radiomics. First order interquartile-range, entropy, uniformity, and GLCM features were found to be statistically significant (p< 0.01). After adjusted for age, gender, and time interval between MRI, IMRT patients developed significant more changes in white matter of the temporal lobe (p=0.03) and corpus callosum (p=0.03) when compared to proton patients.

Conclusion

By employing artificial intelligence and computational methods, we are able to detect global disruption in architecture, connectivity, and function in the brain of head and neck cancer survivors. IMRT affects the brain differently than proton.