The impact of brain computed tomography metadata on the performance of diagnostic artificial intelligence services: a pilot study

The aim of the study is to evaluate the impact of metadata added to the training dataset on the performance of an artificial intelligence system aimed at working with diagnostic brain images. Material and Methods. An expanded set of computed tomography scans of the brain with and without signs of intracranial haemorrhage, supplemented with clinical and technical parameters, was used as a basis. From the dataset, 176 studies were selected (106 for training, 70 for testing), which were specially prepared for input into the neural network: they were segmented with the region of interest (the brain) highlighted, normalized, and the background was removed. The ResNet10 neural network architecture, which is capable of analyzing 3D medical images, was used as the underlying neural network. It was combined with another neural network to form the ResNet10_Meta architecture, which, using a One-HotEnconding (OHE) approach, connects other metadata to analyze images directly. Four parameters were selected as metadata: ‘Manufacturer’, ‘SliceThickness’, ‘PatientAge’, and ‘XrayTubeCurrent’. Results. Twelve experiments were conducted to train the neural network with one, two, or all of the additional parameters added alternately. The model with the addition of the ‘XTube’ parameter showed the highest specificity (82.3 %, 95 % confidence interval (95 % CI) [69.6–94.9]), outperforming the Baseline model without additional metadata (specificity 79.4 %, 95 % CI [66.0–92.8]). The ‘XTube + SliceT’ model with the addition of multiple metadata showed comparatively higher sensitivity (69.7 %, 95 % CI [54.5–84.9]) relative to Baseline model. The model with the addition of all metadata ‘All params’ did not show significant improvements. However, all differences found were statistically insignificant ( p > 0.05). Conclusions. Our data demonstrated that there were no statistically significant differences in the performance of a neural network analyzing diagnostic images without or with metadata added to the training dataset. However, this study is pilot and was conducted on a limited sample, so it remains to be seen whether commercially available artificial intelligence tools can be completely insensitive to image specifications.