Combining Synchrotron Imaging with Artificial Intelligence—The Key for Elucidating Bone Multi-Scale Fracture Origin

The detection and diagnosis of early fractures pose a challenging goal due to the complex and intertwined nature of bone structures. In this context, the field of healthcare is embracing deep learning and artificial intelligence as a means to overcome the subjectivity associated with clinicians' analysis of medical images. However, the current utilization of neural networks is primarily focused on the macro-scale of bone structures, limiting their ability to comprehend the initial stages of crack formation. To gain a better understanding of crack occurrence, it is crucial to adopt a multi-scale perspective and investigate the micro-scale, where the presence of a dense network of lacunae could indicate the initiation or deviation site of a crack. Unfortunately, speculations at this scale can only be made with the assistance of high-resolution imaging techniques, which are time-consuming when it comes to analyzing output images.<br/>In this research, we aim to combine an understanding of the mechanisms behind micro-crack propagation with the promising application of convolutional neural networks (CNNs) and coarse-grained simulation for mechanics prediction.<br/>Trabecular samples obtained from healthy and osteoporotic human femoral heads are considered for this study. These samples are subjected to image-guided failure assessment inside a synchrotron, which emits a monochromatic beam at an energy of 25.6 keV, with a resolution of 1.6 µm. For the detection of cracks and lacunae, a CNN is implemented based on the Keras VGG16 built-in model. During the training and validation stages of the CNN, a total of 648 images are utilized, with a split of 20% for validation and 80% for training. In terms of lacunae identification, the input set consists of three different samples, each containing three images.<br/>The designed CNN demonstrates a high level of accuracy in automatically detecting lacunae and micro-cracks at different compression levels. With the baseline setup, the networks achieve accuracy levels exceeding 0.99 for both cracks and lacunae, with a meanIoU validation metric of over 0.87. This result is particularly encouraging, considering the complexity of the scanning procedure and the challenges encountered when micro-cracks coalesce into larger meso-cracks.<br/>The CNN proves successful in identifying cracks at their initial stages. These cracks become visible and progress further as the applied displacement increases. However, in the final compression stages, cracks are difficult to highlight using the implemented algorithm due to the complete collapse of the bone structure. From a diagnostic standpoint, the higher accuracy in detecting the initial stages of cracks is of great interest, as meso- and macro-scale cracks already represent a critical stage for the subject. Regarding lacunae detection, the CNN achieves optimal results, reducing the computational cost associated with manually segmenting these micrometric elliptical features.<br/>Additionally, experimental images are converted into CG models and then LAMMPS is exploited to run fracture test of the trabecular biological material, which architecture is particularly complex and needs for repeating tests for extrapolating the deviation range. Here, we combined this tool with the up-to-date generative imaging methods to massively generate reasonable artificial phase images and test their fracture, so to get a reliable range of the mechanical properties without actually scanning or testing all of them.<br/>The adopted approaches and methodologies hold promise for developing powerful classifiers and recognition systems to study the microdamage initiation and progression and the mechanical response prediction of bone tissue. Ultimately, this paves the way for the application of machine learning in the study of bone micromechanics.