Enhancing Brain-Tumor Imaging Using a Robust Deep-Learning Approach for Noise Removal and Image Clarity

Received: 3 October 2025 | Revised: 26 December 2025 | Accepted: 6 January 2026 | Online: 4 April 2026

Corresponding author: V. H. Shruti

Abstract

Medical imaging plays a crucial role in the diagnosis of brain tumors. However, the presence of noise, such as Gaussian and Rician noise, degrades image quality, affecting diagnostic accuracy. Existing denoising methods struggle to effectively remove noise without sacrificing critical image details, limiting their usefulness in clinical applications. Hence, this study develops a robust denoising model that preserves essential anatomical structures while efficiently removing noise from brain-tumor images. The study proposes a Three-Layer Convolutional Neural Network (TL-CNN) model to denoise brain-tumor images. The proposed TL-CNN model was trained and tested on the Brain Tumor Segmentation (BraTS) 2021 medical imaging dataset, and evaluated using the Peak Signal-to-Noise Ratio (PSNR). The TL-CNN outperformed existing approaches, achieving a PSNR of 39.84 dB, which demonstrates its superior ability to suppress noise while preserving critical anatomical details. Compared with Zero Contrast-Enhanced Magnetic Resonance Imaging (ZeroCEMR) approach variants, including U-Net Reconstruction Encoder (UNetRe), Cross-Modal U-Net Reconstruction Encoder (CUNetRe), and YOLO-based Reconstruction Encoder (YOLORe), the TL-CNN achieved improvements of 20.34%, 13.15%, and 5.07% in PSNR, respectively. Furthermore, in comparison with the Permutate U-Net combined with Principal Component Analysis (PCA), TL-CNN demonstrated an 11.3% improvement in PSNR. Additional evaluations across different noise types and varying noise levels confirmed the robustness of TL-CNN, which consistently delivered superior denoising performance and maintained high structural fidelity under diverse imaging conditions.