Medical image analysis proliferates, especially in complex tasks such as organ classification, pathology, and abnormalities in X-ray images. Recent studies have focused on using threshold filtering before applying machine learning models and have achieved positive results. This study proposes to use a threshold filtering preprocessing method and then apply machine learning architectures such as multilayer perceptron (MLP), fully connected neural network (FC), and shallow convolutional neural network (Shallow CNN) to classify bone X-ray images into fractured or non-fractured. The study investigated and compared the performance of the models on different thresholds and with different input image sizes. The results show an improvement in accuracy (ACC), with the ACC increasing rate from 63.5% to 83.8% on the original data set when applying the threshold filtering method with thresholds from 80 to 255 and using the input image size of 32 \(\,\times \,\) 32 combined with a simple MLP architecture. These findings demonstrate the effectiveness of combining a threshold filtering preprocessing approach with the choice of an appropriate architecture and appropriate hyperparameters, providing strong support for developing diagnostic models of auto disease. This has the potential to improve disease detection in innovative medicine significantly.