Deep learning-based tool affects reproducibility of pes planus radiographic assessment

Abstract

Angle measurement methods for measuring pes planus may lose consistency by errors between observers. If the feature points for angle measurement can be provided in advance with the algorithm developed through the deep learning method, it is thought that the error between the observers can be reduced. A total of 300 weightbearing lateral radiographs were used for the development of the deep learning-based algorithm, and a total of 95 radiographs were collected for the clinical validation test set. Meary angle (MA) and calcaneal pitch (CP) were selected as measurement methods and measured twice by three less-experienced physicians with the algorithm-based tool and twice without. The intra- and inter-observer agreements of MA and CP measures were assessed via intra-class correlation coefficient. In addition, verification of the improvement of measurement performance by the algorithm was performed. Interobserver agreements for MA and CP measurements with algorithm were more improved than without algorithm. As for agreement with reference standard, combining the results of all readers, both MA and CP with algorithm were greater than those without algorithm. The deep learning algorithm tool is expected to improve the reproducibility of radiographic measurements for pes planus, especially by improving inter-observer agreement.