Utility of machine learning for identifying stapes fixation on ultra-high-resolution CT.

PURPOSE

Imaging diagnosis of stapes fixation (SF) is challenging owing to a lack of definite evidence. We developed a comprehensive machine learning (ML) model to identify SF on ultra-high-resolution CT.

MATERIALS AND METHODS

We retrospectively enrolled 109 participants (143 ears) and divided them into the training set (115 ears) and test set (28 ears). Stapes mobility (SF or non-SF) was determined by surgical inspection. In the ML analysis, rectangular regions of interest were placed on consecutive axial slices in the training set. Radiomic features were extracted and fed into the training session. The test set was analyzed using 7 ML models (support vector machine, k nearest neighbor, decision tree, random forest, extra trees, eXtreme Gradient Boosting, and Light Gradient Boosting Machine) and by 2 dedicated neuroradiologists. Diagnostic performance (sensitivity, specificity and accuracy, with surgical findings as the reference) was compared between the radiologists and the optimal ML model by using the McNemar test.

RESULTS

The mean age of the participants was 42.3 ± 17.5 years. The Light Gradient Boosting Machine (LightGBM) model showed the highest sensitivity (0.83), specificity (0.81), accuracy (0.82) and area under the curve (0.88) for detecting SF among the 7 ML models. The neuroradiologists achieved good sensitivities (0.75 and 0.67), moderate-to-good specificities (0.63 and 0.56) and good accuracies (0.68 and 0.61). This model showed no statistical differences with the neuroradiologists (P values 0.289-1.000).

CONCLUSIONS

Compared to the neuroradiologists, the LightGBM model achieved competitive diagnostic performance in identifying SF, and has the potential to be a supportive tool in clinical practice.