Multiclass datasets expand neural network utility: an example on ankle radiographs

Abstract

Purpose Artificial intelligence in computer vision has been increasingly adapted in clinical application since the implementation of neural networks, potentially providing incremental information beyond the mere detection of pathology. As its algorithmic approach propagates input variation, neural networks could be used to identify and evaluate relevant image features. In this study, we introduce a basic dataset structure and demonstrate a pertaining use case. Methods A multidimensional classification of ankle x-rays (n = 1493) rating a variety of features including fracture certainty was used to confirm its usability for separating input variations. We trained a customized neural network on the task of fracture detection using a state-of-the-art preprocessing and training protocol. By grouping the radiographs into subsets according to their image features, the influence of selected features on model performance was evaluated via selective training. Results The models trained on our dataset outperformed most comparable models of current literature with an ROC AUC of 0.943. Excluding ankle x-rays with signs of surgery improved fracture classification performance (AUC 0.955), while limiting the training set to only healthy ankles with and without fracture had no consistent effect. Conclusion Using multiclass datasets and comparing model performance, we were able to demonstrate signs of surgery as a confounding factor, which, following elimination, improved our model. Also eliminating pathologies other than fracture in contrast had no effect on model performance, suggesting a beneficial influence of feature variability for robust model training. Thus, multiclass datasets allow for evaluation of distinct image features, deepening our understanding of pathology imaging.

AI-based X-ray fracture analysis of the distal radius: accuracy between representative classification, detection and segmentation deep learning models for clinical practice

“…For our classification model, we modified a headless imagenet-pretrained Xception-network based on the keras implementation 19 by adding a global average pooling and a drop-out layer (10% drop-out during training) as well as 2 dense layers (128 and 2 units, respectively). Empirically, 20 model performances could be improved by replacing the superfluous two input layers in the third dimension of imagenet-pretrained networks with filtered copies of the original two-dimensional image after application of a brightness-inversion and an adaptive mean thresholding edge-enhancing filter, respectively. We conjecture generally facilitated fracture delineation by the corresponding and contrasting input information as the reason for improvement, similar to the common practice of performing brightness inversion on radiographs when looking for fracture or pleural lines.…”