Anna K. Jerebko PhD, Presenter: Employee, Siemens AG

Gerardo Hermosillo-Valadez PhD, Abstract Co-Author: Employee, Siemens AG

Vikas C. Raykar PhD, Abstract Co-Author: Researcher, Siemens AG

Andreas Fieselmann, Abstract Co-Author: Employee, Siemens AG

Shiras Abdurahman, Abstract Co-Author: Nothing to Disclose

Thomas Mertelmeier MD, Abstract Co-Author: Employee, Siemens AG Stockholder, Siemens AG

The developed algorithm may be useful in the routine evaluation of BT. The autowindowing algorithm for BT is investigational software and is limited by US law to investigational use.

For accurate diagnosis Breast Tomosynthesis (BT) images must be read with appropriate brightness and contrast settings. The differences in breast density and composition (amount of fatty and glandular tissue, calcifications and scarring) and high inter-observer variability in preferences for contrast and brightness make it difficult to define settings that are optimal for all patients and readers. We developed an algorithm that automatically determines the optimal settings based on radiologist preferences and characteristics of each breast. The algorithm computes a set of breast composition features and uses linear regression to map them to the optimal settings.

The algorithm successfully passed Turing test (p<0.05) according to readers 2 and 4. Reader 1 was the most experienced and was able to identify the expert settings in >50% of the cases. All the readers confirmed that in all cases the computed settings were acceptable.

We designed a Turing test (a test of a machine's ability to exhibit intelligent behavior) to evaluate the algorithm. If a human judge cannot reliably tell the algorithm from the human, the algorithm is said to have passed the test. The algorithm was trained on a set of 28 volumes with optimal settings provided by an expert. Four readers evaluated an independent set of 59 volumes. Each reader was shown the same volume with two settings, one provided by the expert (reader 2) and one by the algorithm. Readers had to choose preferred settings but could also state they were equivalent. The readers did not know which setting came from the expert. For each reader, we computed the fraction (F) of volumes where the expert settings were preferred and calculated 95% confidence intervals (CI) and p-values.Reader 1 had an F of 0.58 [34/59] (95% CI=[0.45, 0.70], p=0.882). Reader 2 had an F of 0.34 [20/59] (95% CI=[0.22, 0.46], p=0.004). Reader 3 had an F of 0.43 [25/59] (95% CI=[0.30, 0.55], p=0.118). Reader 4 had an F of 0.32 [19/59] (95% CI=[0.20, 0.44], p=0.002).

Jerebko, A, Hermosillo-Valadez, G, Raykar, V, Fieselmann, A, Abdurahman, S, Mertelmeier, T, Can a Machine Learning-based Windowing Algorithm Pass Turing Test?.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12031968.html