Keywords:
Not applicable, Quality assurance, Image verification, Perception image, Computer Applications-General, Digital radiography, Lung, Artificial Intelligence, Artificial Intelligence and Machine Learning
Authors:
J. von Berg, D. Bystrov, A. Gooßen, S. Krönke, M. Brück, T. Harder, N. Wieberneit, S. Young; Hamburg/DE
DOI:
10.26044/ecr2020/C-05625
Methods and materials
Artificial intellicence (a combination of convolutional neural networks and probabilistic anatomical atlases) is used to detect anatomical features robustly and consistently even in images not showing diagnostic quality: The convolutional neural networks are trained to detect the lung field border, the posterior ribs, the diaphragm, the clavicle heads, and the spinous processes. Probabilistic anatomical atlases are used to relate these anatomies to each other, to extrapolate incomplete data, and to reject inconsistant false detections (see fig 2).
Based on these robustly detected landmarks several quality figures are geometrically derived that quantitatively represent different quality aspects on patient positioning. Three such quality checkers are realized:
(i) The field-of-view checker measures the four distances between lung field and image borders (collimation) - see fig 3
(ii) The rotation checker balances the distances of both clavicles to the spinous process line. Assymetry α=(dL-dR)/(dL+dR) - see fig 4
(c) The inhalation checker counts the posterior ribs located above the diaphragm - see fig 5
The modules were trained and evaluated with more than 1000 annotated images from several institutions and x-ray system types.
Number of annotated anatomical features:
anatomy
|
training
|
testing
|
total
|
lung borders
|
857
|
211
|
1168
|
clavicles
spinous proc
|
1127
716
|
422
333
|
1549
1049
|
posterior ribs
diaphragm
|
992
1148
|
389
423
|
1381
1571
|