In the realm of low-dose X-ray CT denoising, a plethora of deep learning methodologies have been proposed. However, the prevalent encoder-decoder network paradigms exhibit limitations in their capacity to capture and reconstruct the image’s multi-scale features comprehensively. The commonly used Mean Squared Error (MSE) loss function is often implicated in the over-smoothing phenomena, resulting in the loss of edges and fine details in denoised images. This paper presents a novel Tripartite-Branching Encoder-Decoder Convolutional Neural Network framework (TBED-CNN) for denoising LDCT images. The TBED-CNN framework is characterized by a tripartite processing pathway, each branch specially designed to accentuate multi-scale feature representation, complemented by a loss function that is acutely sensitive to edge preservation. The first branch employs dilated convolutions to capture context expansively, the second utilizes max-pooling to highlight important features, and the third branch integrates an attention mechanism to focus on pivotal pixel-level details. Furthermore, the amalgamation of MSE and gradient loss functions serves to effectively reduces noise while preserving image clarity and structural integrity. Extensive experiments on AAPM low-dose CT dataset demonstrate that TBED-CNN outperforms state-of-the-art models.