Adult Age Estimation of Forensic Pelvis Based on Opitimized Swin Transformer

In recent years, deep learning has been applied to various fields, such as computer vision, pattern recognition etc. Accurate age estimation is essential in anthropology, archaeology and forensic science. Traditional age estimation is generally achieved by the grading and assignment of the bone sites, which is easily affected by the subjective factors of observers. Deep learning has made great progress in medical imaging disease diagnosis and image segmentation, making it possible to observe multiple sites of forensic pelvis simultaneously and being more objective and efficient for adult age estimation. In this paper, a Optimized Swin Transformer approach (OST) is proposed. A database of 164 samples is constructed, composed of MSCT image data and 3D image data firstly. Next, MSCT image data are preprocessed into a uniform size respectively, and are used to expand the number of samples by using data augmentation technique. Then an Optimized Swin Transformer model is constructed with adding a mish active function, a linear layer and sigmoid active function in the end of Swin Transformer model for enhancing the model to learn the complex relationship of pelvic data and representing the extracted features to being mapped into the specified range to generate a probabilistic prediction of bone age. AdamW is used by the optimizer with a segmented learning rate strategy, and MSELoss combined with L1Loss, regarded as the loss function, is utilized to optimize the proposed model. After that, the transfer learning technology is used to train and optimize Swin Transformer model with pre-training parameters. The optimized Swin Transformer model is applied to adult age estimation. Finally, our proposed OST approach is implemented on our constructed database by obtaining better results with coefficient of determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ) being 0.9371, and Root Mean Squared Error (RMSE) being 2.066 . The effectiveness of our proposed OST approach has been fully verified by numerical experimental results.