Medical Image Classification Using Hybrid Deep Learning Methods and Advanced Preprocessing Using Wrist X-ray Images

Abstract

The most common kind of crack, and one that happens rather often, is a wrist fracture. Even though X-ray medical imaging is useful for identifying wrist fractures, the depiction of these breaks may sometimes provide problems. These days, there's a lot of hope that artificial intelligence (AI) might help orthopaedic X-ray interpreters make better, faster fracture diagnoses. The purpose of this investigation is to use “deep learning” (DL) to assist physicians, particularly those who work in emergency departments, in diagnosing wrist X-ray fractures. To build a reliable and effective system that classifies wrist X-ray abnormalities, this study used a MURA (XR_WRIST) dataset from the Stanford Machine Learning Group website. To obtain higher accuracy, the data was preprocessed using various key steps, initially resizing the images and converting them to greyscale; CLAHE was implemented. To mitigate minor noise, Gaussian blur is implemented, and the "haar" wavelet is employed in the discrete wavelet transform (DWT) at level 1 to further reduce noise by zeroing high-frequency coefficients. The model's robustness is ensured by using a 3-fold cross-validation strategy. For classification purposes, the VGG19_AlexNet_AdaptiveHybrid_CNN model is subsequently implemented. Model performance has been assessed using metrics such as the confusion matrix, “Cohen's kappa”(CK), “F1-score”(f-measure), “sensitivity”(Sn), “specificity”(Sp), “recall”(recl), “accuracy”(acc), “precision”(prec), and ROC-AUC. The comparative results of existing and proposed DL models demonstrate thatproposed VGG19_AlexNet_AdaptiveHybrid_CNN model has the highest accuracy99.81%, precision99.9%, recall99.81% and f1-score99.86%. These outcomes show that a proposed model is highly reliable in wrist X-ray abnormalities detection.