Region-based feature enhancement using channel-wise attention for classification of breast histopathological images

• Published in: Neural computing & applications 2023-03, Vol.35 (8), p.5839-5854
• Main Authors: Rashmi, R., Prasad, Keerthana, Udupa, Chethana Babu K.
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Color; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Datasets; Evaluation; Feature extraction; Heterogeneity; Image analysis; Image enhancement; Image Processing and Computer Vision; Image resolution; Machine learning; Nuclei; Original Article; Probability and Statistics in Computer Science; Tumors
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-022-07966-z

Breast histopathological image analysis at 400x magnification is essential for the determination of malignant breast tumours. But manual analysis of these images is tedious, subjective, error-prone and requires domain knowledge. To this end, computer-aided tools are gaining much attention in the recent past as it aids pathologists and save time. Furthermore, advances in computational power have leveraged the usage of computer tools. Yet, usage of computer-aided tools to analyse these images is challenging due to various reasons such as heterogeneity of malignant tumours, colour variations and presence of artefacts. Moreover, these images are captured at high resolutions which pose a major challenge to designing deep learning models as it demands high computational requirements. In this context, the present work proposes a new approach to efficiently and effectively extract features from these high-resolution images. In addition, at 400x magnification, the characteristics and structure of nuclei play a prominent role in the decision of malignancy. In this regard, the study introduces a novel CNN architecture called as CWA-Net that uses a colour channel attention module to enhance the features of the potential regions of interest such as nuclei. The developed model is qualitatively and quantitatively evaluated on private and public datasets and achieved an accuracy of 0.95% and 0.96%, respectively. The experimental evaluation demonstrates that the proposed method outperforms state-of-the-art methods on both datasets.