| Autor: |
Djouima, Hossena, Zitouni, Athmane, Megherbi, Ahmed Chaouki, Sbaa, Salim |
| Předmět: |
|
| Zdroj: |
Revue d'Intelligence Artificielle; Aug2024, Vol. 38 Issue 4, p1097-1107, 11p |
| Abstrakt: |
Automated diagnosis and evolving CNN architectures are improving diagnostic quality in digital breast cancer histopathology images. The study predominantly focuses on classifying the histopathological images of the BreakHis breast cancer dataset into distinct categories: benign and malignant. A primary challenge in this task is the uneven class distribution and limited training samples, which introduce bias and compromise the model's non-malignant classification accuracy. The study utilizes wavelet decomposition on benign images to address class imbalance and enhance the model's ability to accurately classify breast cancer histopathological images. This technique begins by filtering the image with high-pass and low-pass filters, followed by downsampling. The process is then repeated to generate four images representing different components of the original image, enabling precise localization of essential features and denoising. The DenseNet201 convolutional network is chosen for image classification due to its efficiency and accuracy. Our proposal involves concatenating features extracted from specific blocks of the pre-trained DenseNet201 model: pool3_pool, pool4_pool, and conv5_block32_conca. The proposed framework achieves an impressive overall accuracy in classifying both benign and malignant images, maintaining high accuracy rates of 99% in both multi-scale and magnification-independent classifications. These promising results indicate the potential clinical application of this approach in diagnosing diseases. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
