An Adaptive Tissue Characterisation Network for Model-Free Visualisation of Dynamic Contrast-Enhanced Magnetic Resonance Image Data

Dynamic contrast-enhanced magnetic resonanceimaging (DCE-MRI) has become an important source of informationto aid cancer diagnosis. Nevertheless, due to the multi-temporalnature of the three-dimensional volume data obtained fromDCE-MRI, evaluation of the image data is a challenging task andtools are required to support the human expert. We investigatean approach for automatic localization and characterization ofsuspicious lesions in DCE-MRI data. It applies an artificial neuralnetwork (ANN) architecture which combines unsupervised andsupervised techniques for voxel-by-voxel classification of temporalkinetic signals. The algorithm is easy to implement, allows forfast training and application even for huge data sets and canbe directly used to augment the display of DCE-MRI data. Todemonstrate that the system provides a reasonable assessment ofkinetic signals, the outcome is compared with the results obtainedfrom the model-based three-time-points (3TP) technique whichrepresents a clinical standard protocol for analysing breast cancerlesions. The evaluation based on the DCE-MRI data of 12 casesindicates that, although the ANN is trained with impreciselylabeled data, the approach leads to an outcome conforming with3TP without presupposing an explicit model of the underlyingphysiological process.