Bone microstructure identification using deep learning: A pilot study

Introduction: Several methods can be used for age estimation during forensic identification. Bone histology remains one of the few options when bones are highly fragmented or affected taphonomically. However, the method presents limitations, notably the substantial training and expertise needed to ensure reliable results. Our pilot study explored the potential of deep learning to automatically identify osteonal structures.
Materials and Methods: Two observers (JGGD, RH) manually outlined the contours of 250 intact osteonal structures following protocols from the literature, in 70 rib thin-section (0.2mm) microphotographs (10X magnification, 2592x1944 pixels). Fragmentary osteons were also annotated as a resource for use in future work. A U-net, a deep learning architecture well-suited to medical image segmentation, was trained to segment intact osteons, i.e., to perform binary labelling of pixels as belonging to intact osteons or otherwise. Fifty images were used for development and 20 for testing. A Keras implementation of a U-net was trained with binary cross-entropy loss, Adam optimization, and data augmentation.
Results: U-net segmentation of test images gave an F1 score (Dice coefficient) of 0.67. This initial result is encouraging given the visually complex nature of these structures, and the confounding presence of fragmented osteons.
Conclusions: The present project offers a preliminary investigation on the identification of osteonal structures in human cortical bone using machine learning. We anticipate that performance will be improved through expansion of the training dataset as well as refinement of the segmentation model. Future work will build on this pilot study to explore automated osteon counting, and segmentation of individual osteon instances, including fragmented osteons, with the potential of providing a standardization protocol on bone microstructure analysis and a wider application of histology for identification purposes.