Open Access

ARTICLE

Classifying Hematoxylin and Eosin Images Using a Super-Resolution Segmentor and a Deep Ensemble Classifier

P. Sabitha^*, G. Meeragandhi

Department of Computer Science and Engineering, Sathyabama Institute of Science and Technology, Chennai, Tamilnadu, India

* Corresponding Author: P. Sabitha. Email:

Intelligent Automation & Soft Computing 2023, 37(2), 1983-2000. https://doi.org/10.32604/iasc.2023.034402

Received 16 July 2022; Accepted 23 November 2022; Issue published 21 June 2023

Abstract

Developing an automatic and credible diagnostic system to analyze the type, stage, and level of the liver cancer from Hematoxylin and Eosin (H&E) images is a very challenging and time-consuming endeavor, even for experienced pathologists, due to the non-uniform illumination and artifacts. Albeit several Machine Learning (ML) and Deep Learning (DL) approaches are employed to increase the performance of automatic liver cancer diagnostic systems, the classification accuracy of these systems still needs significant improvement to satisfy the real-time requirement of the diagnostic situations. In this work, we present a new Ensemble Classifier (hereafter called ECNet) to classify the H&E stained liver histopathology images effectively. The proposed model employs a Dropout Extreme Learning Machine (DrpXLM) and the Enhanced Convolutional Block Attention Modules (ECBAM) based residual network. ECNet applies Voting Mechanism (VM) to integrate the decisions of individual classifiers using the average of probabilities rule. Initially, the nuclei regions in the H&E stain are segmented through Super-resolution Convolutional Networks (SrCN), and then these regions are fed into the ensemble DL network for classification. The effectiveness of the proposed model is carefully studied on real-world datasets. The results of our meticulous experiments on the Kasturba Medical College (KMC) liver dataset reveal that the proposed ECNet significantly outperforms other existing classification networks with better accuracy, sensitivity, specificity, precision, and Jaccard Similarity Score (JSS) of 96.5%, 99.4%, 89.7%, 95.7%, and 95.2%, respectively. We obtain similar results from ECNet when applied to The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) dataset regarding accuracy (96.3%), sensitivity (97.5%), specificity (93.2%), precision (97.5%), and JSS (95.1%). More importantly, the proposed ECNet system consumes only 12.22 s for training and 1.24 s for testing. Also, we carry out the Wilcoxon statistical test to determine whether the ECNet provides a considerable improvement with respect to evaluation metrics or not. From extensive empirical analysis, we can conclude that our ECNet is the better liver cancer diagnostic model related to state-of-the-art classifiers.

---

Keywords

---