Convolutional Neural Networks for Automatic Risser Stage Assessment

Abstract

A deep learning network was developed to determine Risser stage on adolescent pelvic radiographs. The network had similar accuracy to expert readers, and thus could be implemented to aid physicians to provide a second opinion on staging. Key PointsThe developed deep learning method to automate Risser stage assessment reached 78.0% accuracy, which was comparable to 74.5% agreement between expert readers. Risser stage assessment using deep learning models is promising for the evaluation of skeletal maturity in AIS and could reduce the propagation of error biases within clinical files.

Automatic bone maturity grading from EOS radiographs in Adolescent Idiopathic Scoliosis

“…First, the results obtained show a significant improvement of our ResNet101-SVM compared against the GLCM-SVM and the VGG16 model [23]. The previous study using VGG16 model is built with 1830 pelvic radiographs and trained in 8 hours [23]. Comparatively, ResNet101-SVM is built with 257 EOS radiographs and trained in one quarter of hour.…”

“…Some studies on measurement variability have shown acceptable results [15,14]. However, the latest studies are less optimistic and consider agreement moderate [37,23]. While the formal definition of the Risser sign relies on the observation of the iliac crests, other centers of ossification can be explored for proper grading.…”

“…While the formal definition of the Risser sign relies on the observation of the iliac crests, other centers of ossification can be explored for proper grading. Thus, Kaddioui et al [23] demonstrated that the entire pelvic region could be used to better determine the Risser sign. Besides, several metrics make it possible to estimate bone age by observing other regions of the body [2].…”

Artificial intelligence development in pediatric body magnetic resonance imaging: best ideas to adapt from adults

“…For Risser stage classification, in 1,830 radiographs performed for adolescent idiopathic scoliosis, the CNN was comparable to slightly better than six expert graders, with 78.0% versus 74.5% accuracy, respectively. Additionally, the kappa coefficient for this CNN was 0.72, which exceeded the 0.65 for the human graders (kappa coefficient measures inter-rater agreement; 0.72 indicates substantial agreement) [ 9 ].…”

Investigating the Impact of Explanation on Repairing Trust in Ai Diagnostic Systems for Re-Diagnosis

“…Advanced forms of machine learning developed in the late 1990s which made 14 computational devices more accurate and robust. Thus, the clinical application of AI has been most rapid in image-intensive fields such as radiology, radiotherapy, pathology, dermatology, ophthalmology and image-guided surgery (Codella et al, 2018;Dominic et al, 2019;Esteva et al, 2017;Jha and Topol, 2016;Kaddioui et al, 2020;Kundu et al, 2017;Naylor, 2018;Piccini et al, 2020;Ting et al, 2019). Most of these AI systems have high accuracy in prediction and can make very fast diagnoses that make them capable of making the whole healthcare system faster than before.…”