TL;DR
This systematic review synthesizes 39 recent studies on medical image enhancement, highlighting persistent quality issues like low contrast and noise across various modalities. It reveals a significant shift towards deep learning and AI-driven approaches, while also underscoring the need for standardized evaluation metrics and more diverse datasets to improve diagnostic accuracy.
What Problem Does This Paper Address?
Medical imaging is indispensable for diagnosis, but images often suffer from inherent quality issues such as noise, artifacts, and low contrast. These imperfections can obscure critical details, hindering accurate diagnosis and effective treatment planning. While imaging technologies have advanced, the challenge of enhancing image quality to maximize diagnostic potential remains. This paper addresses the need for a comprehensive overview of the current landscape of medical image enhancement, covering the challenges, the latest advancements in techniques, and the crucial evaluation metrics used to assess their effectiveness. It aims to bridge the gap left by existing reviews that often focus on specific modalities or techniques, providing a holistic view of the field.
Methods
The authors conducted a systematic literature review following the PRISMA guidelines. They searched two major databases (ScienceDirect and Web of Science) for studies published within the last five years (2020-2025) using specific keywords related to medical image enhancement and quality issues. Eligibility criteria were applied to include studies focusing on medical imaging, employing enhancement techniques (traditional, deep learning, or hybrid), and providing quantitative or qualitative evaluations. Exclusion criteria removed irrelevant studies like reviews, books, and those without empirical validation. Data extraction covered authors, image quality issues, modalities, datasets, enhancement methods, software, evaluation metrics, and outcomes. A quality assessment framework was used to evaluate the credibility of the selected articles.
Datasets
The paper discusses various datasets used across the reviewed studies, but does not introduce any new benchmarks. Examples of datasets mentioned include:
- MIAS, LITFL (for mammograms, X-rays, CT scans)
- Radiology Assistant, MR-TIP, BrainWeb (for MRI)
- University Hospital ‘Policlinico G. Martino’ dataset (for chest X-rays)
- MedPix, OpenI, MRTIP (for various modalities)
- Guangzhou Dataset (for X-ray)
- Handheld Device and High-End Device, Beijing Institute of Ophthalmology, Tongren Hospital private dataset (for retinal images)
- COVID-19 CT and X-ray image datasets (e.g., from GitHub, SIRM)
- fastMRI, NYU, ChestX-ray, Kvasir, CVC-ClinicDB, ETIS-Larib Polyp DB, STARE, DRIVE, REFUGE, CORN-2, BraTS 2018, TCGA-LGG, etc.
Key Insights and Ideas
- Low contrast and noise are persistent villains: These two issues are the most frequently reported challenges across a wide range of medical imaging modalities, significantly impacting diagnostic accuracy.
- Deep learning is taking over: There's a clear trend moving away from purely conventional mathematical methods towards AI-driven and deep learning-based approaches, particularly CNNs and GANs, which demonstrate superior performance.
- Hybrid approaches show promise: Combining traditional and deep learning techniques offers a balanced approach, potentially leveraging the interpretability of conventional methods with the power of AI.
- Specialized modalities are underexplored: While MRI and multi-modal imaging receive significant attention, areas like bone scintigraphy, endoscopy, and histopathology are relatively under-researched, despite their unique challenges.
- Standardization is a critical bottleneck: A lack of standardized evaluation metrics and benchmarking datasets hinders reproducibility and comparative analysis across studies, making it difficult to truly assess the progress in the field.
- Non-reference-based IQA metrics are gaining traction: As obtaining ground truth references is often impractical in medical imaging, there's a growing reliance on metrics that can assess image quality without a pristine reference.
Key Results
The review analyzed 39 studies, identifying the following:
- Image Quality Issues: Low contrast (33 occurrences), noise (15), brightness inconsistencies (7), uneven illumination (8), blurring (5), artifacts (3), and color imbalance (2) were the most prevalent issues.
- Enhancement Techniques: 29 studies used conventional methods, 9 focused on deep learning, and 1 used a hybrid approach.
- Image Quality Assessment (IQA) Metrics: A total of 65 IQA metrics were introduced, with a significant majority (42) being non-reference-based. 18 studies used both reference-based and non-reference-based metrics, 9 used only reference-based, and 12 used only non-reference-based.
- Quality of Reviewed Articles: The reviewed articles were generally of high quality, with scores ranging from 75% to 95%, indicating robust research designs and detailed results. 31 out of 39 studies scored 85% or above.
Limitations and Future Work
Limitations: * Lack of Standardization: A critical limitation is the absence of standardized software usage and benchmarking across studies, hindering reproducibility and comparative analysis. Many studies omit software details. * Dataset Scarcity: Deep learning models often require large, labeled datasets, which are scarce in medical imaging. Existing datasets may not fully capture real-world clinical variability. * Limited Generalizability: Some methods show limited generalizability to other modalities or are untested on diverse datasets. * Computational Cost: Many advanced methods, especially deep learning ones, are computationally intensive, posing challenges for real-time deployment. * Underexplored Modalities: Specialized modalities like bone scintigraphy, endoscopy, and histopathology require more research attention.
Future Work: * Adaptive Algorithms: Develop adaptive algorithms for contrast enhancement and noise reduction, especially for underexplored modalities. * GANs for Histology: Utilize Generative Adversarial Networks (GANs) to improve staining normalization in histology. * Open-Access Datasets & Benchmarks: Create open-access datasets reflecting real-world variability and standardized image quality benchmarks. * Cross-Modality Preprocessing: Develop solutions for unified AI-driven enhancements across different imaging domains. * Explainable AI (XAI): Integrate XAI to increase transparency in automated image processing, particularly for cancer detection. * Standardized Protocols: Establish standardized imaging protocols across institutions for diagnostic consistency. * Self-supervised/Unsupervised Learning: Explore these models to mitigate the scarcity of labeled medical imaging data. * Hybrid Approaches: Focus on hybrid methods that balance enhancement with structural preservation to avoid artifacts or degradation of diagnostic details. * Lightweight AI Models: Prioritize lightweight AI models optimized for real-time edge-device processing. * Standardized IQA Metrics: Develop standardized IQA metrics and evaluation protocols to enhance reproducibility and applicability. * Deep Learning IQA: Integrate statistical and deep learning-based perceptual assessments for IQA, and develop large-scale medical IQA datasets for better training.
Summary updated