Breast cancer anomaly detection based on the possibility theory with a clustering paradigm. (January 2023)

• Record Type: Journal Article
• Authors: Elleuch, Jihen Frikha; Mehdi, Mouna Zouari; Belaaj, Majd; Benayed, Norhène Gargouri; Sellami, Dorra; Damak, Alima
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: Possibility theory -- Clustering -- Gaussian mixture model -- Mass -- Microcalcification; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104043
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Breast cancer early diagnosis is a major concern for reducing deadly cases. Related breast tissue anomalies are masses and micro-calcifications, which characterization is essential to obtain a correct diagnosis. At this step, high false negative rates are observed, revealing a high ambiguity, which are crucial to handle with conventional approaches. Possibility theory offers a powerful paradigm enabling to handle a high uncertainty level. For enhancing the classification's accuracy in case of scattered classes, we have redefined our classes based on clustering. Under this paradigm, we adopt in the clustering optimization a criteria based on the Dubois and Prade possibility transformation. In this research, we propose a new possibilistic modeling where a clustering based aggregation of features is applied followed by a fusion of the possibility distributions, for decision making. The above strategy has been applied on two case studies in breast cancer screening: mass and micro-calcification detection, where a set of challenging features are considered. Validation on the CBIS-DDSM public dataset has been undertaken. The proposed system, yielding a mass and micro-calcification detection accuracy respectively of 95.4% and 99.4% (at a specificity respectively of 96% and 100% and with an AUC respectively of 0.96 and 0.996), can be used to assist medical practitioners. Highlights: Development of general formalism for possibility transformation of medium level features and their fusion, using the asymmetric Dubois and Prade transformation. Optimization of the clustering parameters (number of adequate clusters for each feature) based on a new related metric, which is the AUC taking as a confidence degree level the possibility consistency measure. High rates of breast tissues anomaly detection accuracy: for the mass detection, an accuracy rate of 95.4% is obtained, while an accuracy rate of 99.4% is obtained for the micro-calcification detection.