A Comparative Study of Chest X-Ray Image Enhancement Techniques for Pneumonia Recognition
| Autor: | Sabrina Nefoussi, Idir Amine Amarouche, Abdenour Amamra |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Computer science
business.industry media_common.quotation_subject 020208 electrical & electronic engineering 0206 medical engineering Pattern recognition 02 engineering and technology Image enhancement medicine.disease 020601 biomedical engineering Pneumonia 0202 electrical engineering electronic engineering information engineering X ray image medicine Preprocessor Contrast (vision) Adaptive histogram equalization Artificial intelligence business Histogram equalization Unsharp masking media_common |
| Zdroj: | Advances in Computing Systems and Applications ISBN: 9783030694173 CSA |
| DOI: | 10.1007/978-3-030-69418-0_25 |
| Popis: | Unlike the identification of subjects with salient features in natural images, the visual similarity between the pathological features of Chest X-Ray (CXR) images complicates the distinction and interpretation of pneumonia signs. In this work, we aim to enhance Chest X-ray images by applying several image pre-processing techniques such as histogram equalization, contrast limited adaptive histogram equalization (CLAHE) and Unsharp mask. For instance, the fact of being the most performant natural image enhancement algorithms raises the question of whether they can achieve similar performance on CXR imagery. Hence, our objective is to investigate these enhancement techniques for the task of Pneumonia classification with deep neural architectures. To validate our findings, we provide a comparative study on the largest public pneumonia dataset of the Radiological Society of North America (RSNA) dubbed “Pneumonia Detection Challenge Dataset”. The performance was measured in the two cases of balanced and imbalanced positive (pneumonia) and negative (no pneumonia) classes, where the preprocessing is intended to mitigate the bias toward the dominant class. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|