DEEP LEARNING-DRIVEN IMAGE ENHANCEMENT AND LESION SEGMENTATION FOR IMPROVED DIAGNOSTIC ACCURACY IN NOISY MEDICAL IMAGES
Abstract
Medical imaging technologies, including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and ultrasound, play a significant role in disease diagnosis and treatment planning. Therefore, medical images are sometimes affected by noise, low contrast, motion artifacts, and intensity variations that reduce image quality and complicate lesion detection. This study proposes a deep learning-based framework for image enhancement and lesion segmentation to improve diagnostic accuracy in noisy medical images. The proposed framework adopts a convolutional autoencoder-based image enhancement model with a lightweight U-Net segmentation architecture for creating an effective medical image analysis pipeline. The experimental study was conducted using liver CT images from the Liver Tumor Segmentation (LiTS) dataset. In this context, preprocessing operations such as resizing, normalization, lesion-mask extraction, and Gaussian noise simulation have been implemented to generate noisy medical imaging scenarios. The autoencoder improvement module has been designed to suppress noise and reconstruct improved images while preserving significant anatomical structures. The improved images were used for automated lesion segmentation utilizing the U-Net-based framework. Performance evaluation was carried out using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Dice Score, and Intersection over Union (IoU). The PSNR of 24.57 dB and the SSIM of 0.792 indicate that the enhancement framework can reconstruct high-quality images while retaining their structural information. It has been found that the Dice Score and IoU Score of the segmentation framework are 0.134 and 0.072, respectively, under noisy imaging conditions. Comparative analysis revealed that lesion localization and segmentation consistency were enhanced by enhancement-assisted segmentation compared with direct segmentation of noisy images. The study also demonstrated that the cross-modality analysis is adaptable to both MRI-like and ultrasound-like simulated images. The results show that combining image enhancement and lesion segmentation is a promising approach to achieve more reliable and efficient noisy medical image analysis. The proposed framework offers a promising basis for developing future computer-aided diagnostic systems and automated healthcare applications for noisy medical imaging environments.