<p>Liver tumor segmentation and diagnosis from medical images play an important role in treatment planning. However, the conventional process of detecting and segmenting liver tumors is a time-consuming and highly observer-dependent task. Therefore this study develops a multi-scale transformer-based neural network method called TransDense121-UNet for accurate detection and segmentation of liver tumors. This method includes data preprocessing, detection, and segmentation phases. The proposed method is evaluated on various datasets such as 3DIRCADb, LiTS2017, and liver tumor segmentation dataset. Multi-scale Axis Attention is used to extract various features of different scales including intensity, size, shape, and location during liver tumor segmentation. This model applies a multi-path framework that combines multi-scale atrous convolutions in parallel with an axis attention layer. It increases effectiveness as retains a broad receptive field at various dilation rates and provides enhanced features without compromising resolution, resulting in the accurate detection of small tumors. In contrast, the axis attention mechanism is used to capture long-range dependencies for better integration of contextual information which is useful in segmenting complex liver images. In the detection and segmentation phase, TransDense121-UNet is employed which encompasses utilized UNet, DenseNet121, and Enhanced Transformer. By leveraging the unique capabilities of dense block relevant features are extracted, while robust enhanced transformer systems address issues of computational complexity. Through the experimental validation, the proposed approach achieved an accuracy of 98.78% and a Dice coefficient of 0.989, demonstrating its effectiveness in precisely detecting and segmenting liver tumors.</p>