Research Output
A proposed framework for improving the detection and classification of Ki67 expression in Astrocytoma histopathological images
Feature Targeted Image Enhancement for Acute Myeloid Leukemia
1 - 2 of 2
No Thumbnail Available
Publication
Detecting and classifying the Ki67 expression is a challenging process. The inconsistency in staining intensity and the variations in image quality are the main factors that may reduce the performance of an automated system. Therefore, this study proposes a framework that objectively improves detecting and classifying Ki67 expression in astrocytoma histopathological images. The proposed framework began with implementing the double stain normalization procedure to reduce the colour-staining intensity variations. Then, the system automatically enhanced the morphological features of the Ki67 expression. The following step was to segment the enhanced images by using the U-Net network model. The last step of the proposed framework was to localize and classify the Ki67 expression based on the modified YOLOv3 model. In conclusion, the proposed YOLOv3 model produced a high detection result with a mean average precision of 0.80 for detecting Ki67-positive cells and 0.87 for detecting Ki67-negative cells.
Publication
Image enhancement is one of the pre-processing steps in various computer vision applications. The current image enhancement algorithm typically applies uniform enhancements across the entire image where this approach often falls short of accurately highlighting or enhancing the specific features due to the influence of the background color. Therefore, this paper proposes a feature-targeted image enhancement technique. Feature-targeted image enhancement (FTIE) algorithm is the improvement over the conventional technique. This method will only enhance the targeted feature instead of the entire image. Therefore, the targeted feature will be enhanced accurately without the influence of the background image. The FTIE method was done by extracting the target feature from the original images and then applying the enhancement method to that region only. Based on the 80 acute myeloid leukemia images, the proposed method showed a promising result, where the comparative analysis shows that the image produced from the proposed method surpasses other conventional methods in terms of structural similarity index (0.995), universal image quality index (0.996), peak signal-to-noise ratio (30.803), mean absolute error (0.002), correlation coefficient (0.997) and contrast enhancement-based image quality (1.743) values.