[I254] From image quality to care outcome

Abstract Medical physicists have a long tradition of measuring image quality with objective metrics including contrast, noise and resolution, and their frequency-based derivatives. These methods have supported our main tasks related to quality assurance and optimisation. Along with the technical imaging modality development, the optimisation process has transformed into more demanding and multi-professional challenge where the image quality metrics should evolve accordingly from technical towards clinical presentation. In the parameter level, this development may include clinical task function and observer related parameters supplementing the traditional MTF and NPS parameters. New methods enable model observer based detectability and diagnostic accuracy estimates. Ultimately, we should aim beyond the concept of technical quality, to extend our methods and knowledge towards measuring and optimising the diagnostic value in terms of care outcome. Modern radiological imaging technology, reconstruction and post-processing techniques provide new and mostly non-linear image output. In part, this explains the need for more complicated analysis of image quality compared to the traditional and more linear image output. Improvement in radiological optimisation requires also patient-specific and indication-specific adjustment of imaging parameters and analysis methods. One size and purpose simply does not fit all patients and applications. Both of these aspects – nonlinearity and patient/indication specificity – aim to improve diagnostic information content and representation of indication-specific image features in radiology. Improved optimisation process and more consistent imaging quality (evaluated by target value, its uncertainty and precision) require objective and quantitative connections from diagnostic and technical parameters to clinical outcome parameters. Comprehensive methodology to enable this approach involves combining several types of data together, as described in a recent publication from an international summit [1] . Artificial intelligence (AI) based deep learning methods – including data quality control and validation – are prerequisites for this kind of data analysis, due to inherent non-linearity of the problem and large amount of heterogeneous data which is not equitable by traditional methods. Our medical physicist professional role should follow this development and incorporate AI & deep learning topics accordingly into our educational programs.