BoneSegNet: Enhanced 2D-TransUnet Model for Multiclass Semantic Segmentation of X-Ray Images of Human Hand Bone

Multiclass segmentation of hand bones in X-ray images is crucial for various medical applications, including diagnostic assistance and surgical planning. This study presents BoneSegNet, a deep learning-based segmentation model designed to classify hand bones into seven distinct anatomical classes: Distal phalanges, Intermediate phalanges, Proximal phalanges, Metacarpals, Carpals, Ulna, and Radius. Our model, BoneSegNet leverages the strengths of both convolutional neural networks (CNNs) and transformer encoders incorporating the UNet architecture. Our model architecture begins with patch and positional embeddings of the input image, followed by a series of transformer encoder layers to capture long-range dependencies and contextual information. The transformer encoder is equipped with multi-head attention mechanisms and multi-layer perceptron’s (MLPs) with GELU activation and dropout for regularization. Skip connections at specific layers are stored and later utilized in the decoder. The decoding path consists of a series of deconvolution and convolution blocks designed to progressively upscale the feature maps while incorporating skip connections from the encoder. This hierarchical structure ensures detailed spatial information is preserved and enhanced. The final output layer employs a 1x1 convolution with a sigmoid activation function to produce the multiclass segmentation map. The model is trained and evaluated on a dataset of hand X-ray images. It achieves an average accuracy of 99.25%, an average Dice score of 0.81, an average mIoU score of 0.70, sensitivity of 0.79, and specificity of 0.997. These results demonstrate the model’s effectiveness in accurately segmenting and classifying hand bones, making it a valuable tool for medical professionals.