Keywords:
Chest, Artificial Intelligence, Radiographers, Thorax, Digital radiography, Technology assessment, Education and training, Quality assurance, Retrospective, Experimental, Performed at one institution
Authors:
S. Young1, M. Kotnik2, L. Lin3, M. Sevenster4, N. Wieberneit1, T. Harder1, S. Krönke1, D. Bystrov1, H. J. Lamb5; 1Hamburg/DE, 2Šmartno pri Slovenj Gradcu/SI, 3Leiden /NL, 4Eindhoven/NL, 5Leiden/NL
DOI:
10.26044/ecr2020/C-12328
Methods and materials
We designed three AI-modules based upon radiological image quality guidelines [1]. These modules are described in more detail in a complementary abstract [2]), and measure the following quality parameters in chest PA radiographs:
- collimation – the distances from the lung field to the image border in all four directions (cranial Δc, abdominal Δa, left Δl and right Δr),
- axial rotation via the symmetry of clavicle head positions w.r.t. medial axis of the thorax, α = (dL – dR) / (dL + dR)
- degree of inhalation via evaluation of posterior ribs superior to diaphragm.
The software measured these parameters in 24,829 consecutive adult images, and for a rater study we selected 1000 cases in three experiments as illustrated in figure 3. Each experiment focused on one of the patient positioning aspects individually (600 cases for FOV, 200 for ROT and 200 for INH). We included cases from across the full range of relevant parameters (images classified as “OK” and “NOT_OK” using thresholds identified in preceding image rater studies [3]) and in all relevant directions of deviation. The most complicated aspect is collimation (FOV) since this may deviate in each of four directions (cranial (N), left (E), abdominal (S) or right (W) using a compass analogy). Furthermore, this deviation may occur in different directions (e.g. collimation may be either “too-narrow” or “too-wide”), and so “NOT_OK” cases were selected from both tails of the distribution. The selection strategy implies that the distribution of presented cases (see figure 3) deviated significantly from the distribution observed in daily clinical routine (however, note also that within a given “experiment”, no specific strategy was used for all other parameters resulting in a random selection from the natural distribution for these parameters).
Two raters (radiologists) used two categories (“good”, “unacceptable”) to assess patient positioning w.r.t. each of the aspects of positioning (field-of-view, rotation and inhalation) as well as providing an overall assessment. Then, concordant expert classifications were compared to the classes predicted by the software for each quality aspect in order to determine sensitivity and specificity.