Artificial Intelligence and Image Analysis for the Identification of Endometrial Malignancies: A Comparative Study

Artificial Intelligence and Image Analysis for the Identification of Endometrial Malignancies: A Comparative Study

Abraham Pouliakis, Vasileia Damaskou, Niki Margari, Efrossyni Karakitsou, Vasilios Pergialiotis, George Valasoulis, George Michail, Charalampos Chrelias, George Chrelias, Vasileios Sioulas, Alina-Roxani Gouloumi, Nektarios Koufopoulos, Martha Nifora, Andriani Zacharatou, Sophia Kalantaridou, Ioannis G. Panayiotides
Copyright: © 2020 |Pages: 37
DOI: 10.4018/978-1-7998-2390-2.ch005
(Individual Chapters)
No Current Special Offers


The aim of this study is to compare machine learning algorithms (MLAs) in the discrimination between benign and malignant endometrial nuclei and lesions. Nuclei characteristics are obtained via image analysis and were measured from liquid-based cytology slides. Four hundred sixteen histologically confirmed patients were involved, 168 healthy, and the remaining with pathological endometrium. Fifty percent of the cases were used to three MLAs: a feedforward artificial neural network (ANN) trained by the backpropagation algorithm, a learning vector quantization (LVQ), and a competitive learning ANN. The outcome of this process was the classification of cell nuclei as benign or malignant. Based on the nuclei classification, an algorithm to classify individual patients was constructed. The sensitivity of the MLAs in training set for nuclei classification was in the range of 77%-84%. Patients' classification had sensitivity in the range of 90%-98%. These findings indicate that MLAs have good performance for the classification of endometrial nuclei and lesions.
Chapter Preview


The endometrium is a human system with inherent complexity because due to the cyclic regeneration during each menstrual cycle under estrogen effects. These effects, as well as, sex steroids, oncogene products, growth factors and peptides (Murphy, Murphy, & Friesen, 1987; Shyamala & Ferenczy, 1981) may cause pre-malignant and malignant transformation of the endometrium. Thus early diagnosis is crucial because it is strongly associated with patient management and on-time therapy. However, and despite the fact that the last decades the incidence of endometrial cancer has increased (Jemal, Siegel, Ward et al., 2006), there is no worldwide and well-established screening method for early detection of endometrial cancer or pre-malignant stages.

The standard procedure for the cytological evaluation of endometrial samples is the cytomorphological evaluation of Papanicolaou-stained direct smears. Despite there is long experience in this technique, there can be errors due to the sampling procedure and due to the remaining cells on the sampling device (Goodman & Hutchinson, 1996) and other errors due to the presence of protein, mucus and blood components (van der Graaf, Vooijs, Gaillard et al., 1987) that may obstruct the cytological view. Such errors can be avoided via liquid-based cytology (LBC). This technique gradually replaces the conventional direct smear preparations. LBC was initially applied for cervical cancer detection and now it is also applied in the assessment of endometrial lesions (Buccoliero, Gheri, Castiglione et al., 2007; Marasinghe, Chintana, Karunananda et al., 2007; Papaefthimiou, Symiakaki, Mentzelopoulou, Tsiveleka et al., 2005). LBC is combined with preparation of single-cell layer slides (mono-layer), thus facilitates the measurement of nuclear features, because cell overlap is reduced in mono-layer specimens. Therefore, an objective cytomorphological discrimination of lesions on the basis of shape and density features is facilitated, especially when such morphological features can be measured by computers.

Machine Learning and in general Artificial Intelligence (AI) techniques are not new in medicine (Almeida & Noble, 2000; Foran, Chen, & Yang, 2011; Grabe, Lahrmann, Pommerencke et al., 2010; Karakitsos, Ioakim-Liossi, Pouliakis et al., 1998; Krzysztof & Krzysztof, 2016; Luo, Ye, Ng et al., 2015; Markopoulos, Karakitsos, Botsoli-Stergiou, Pouliakis, Gogas et al., 1997; Pantazopoulos, Karakitsos, Iokim-Liossi, Pouliakis, Botsoli-Stergiou et al., 1998; Pitoglou, 2018; Salamalekis, Pouliakis, Margari et al., 2019; Seffens, Evans, Minority Health et al., 2015; Siristatidis, Pouliakis, Chrelias et al., 2011; Siristatidis, Vogiatzi, Pouliakis et al., 2016; Su, Xu, He et al., 2016; Vilhena, Vicente, Martins et al., 2017; Vogiatzi, Pouliakis, & Siristatidis, 2019). During the last decades, numerous applications have been reported; these involve either classical statistical models (Cochand-Priollet, Koutroumbas, Megalopoulou et al., 2006; Georgoulakis, Pouliakis, Koutroumbas et al., 2008; Koutroumbas, Pouliakis, Megalopoulou et al., 2006; Megalopoulou, Koutroumbas, Pouliakis et al., 2006; Tzivras, Megalopoulou, Pouliakis et al., 2008) as well as more advanced techniques, such as neural networks. In the field of, oncology-related medical disciplines, and especially for cytopathology there have been reported numerous efforts (Karakitsos, Cochand-Priollet, Guillausseau et al., 1996; Karakitsos, Cochand-Priollet, Pouliakis et al., 1999; Karakitsos, Megalopoulou, Pouliakis et al., 2004; Karakitsos, Pouliakis, Kordalis et al., 2005; Karakitsos, Pouliakis, Koutroumbas et al., 2000; Karakitsos, Stergiou, Pouliakis et al., 1997; Karakitsos, Stergiou, Pouliakis et al., 1996; Markopoulos, Karakitsos, Botsoli-Stergiou, Pouliakis, Ioakim-Liossi et al., 1997; Pantazopoulos, Karakitsos, Iokim-Liossi, Pouliakis, & Dimopoulos, 1998; Pantazopoulos, Karakitsos, Pouliakis et al., 1998). Despite there is some literature related to endometrial cytological material evaluation by MLAs (Karakitsos, Kyroudes, Pouliakis et al., 2002; Pergialiotis, Pouliakis, Parthenis et al., 2018; Abraham Pouliakis, Margari, Margari et al., 2014; A. Pouliakis, Margari, Karakitsou et al., 2018; A. Pouliakis, Margari, Karakitsou et al., 2019; Zygouris, Pouliakis, Margari et al., 2014), to the authors’ knowledge, up to date, publications reporting the comparison of MLAs’ performance is rather poor. This is a well-known issue since there is no common method to report results and moreover the measurements used for the classifications are different.

Key Terms in this Chapter

Machine Learning (ML): It is a subset of artificial intelligence. The main characteristic of this discipline is related to the study of algorithms (including statistical models) that can be used by computerized systems in order to perform learning and recognition tasks however without using extensive and explicit instructions, instead learning of patterns and inference is used.

Artificial Neural Networks (ANNs): They are complex computational models inspired by the human brain nervous system, capable of learning and pattern recognition (an AI-related branch).

Cytopathology: A specialty of medicine relevant to the study and diagnosis of diseases by the examination of cells.

Learning Vector Quantizer (LVQ): Is a pattern-based artificial neural network that belongs to the supervised networks family. A winner-take-all learning approach is applied.

Artificial Intelligence (AI): It is also called machine intelligence. The term is used often to for machines (i.e., computers) that can mimic some cognitive functions of humans, for example learning/recognizing or solving problems.

Complete Chapter List

Search this Book: