Original Research ArticleThe use of the Hellwig's method for feature selection in the detection of myeloma bone destruction based on radiographic images
Introduction
Multiple myeloma (MM) is a haematologic cancer that is characterised by bone marrow infiltration by pathological plasma cells, the presence of abnormal paraprotein in serum and/or urine, and commonly occurring symptoms in the form of bone osteolysis. Despite many studies, scientists have failed to understand the causes of myeloma. The disease develops in the bone marrow, and the lesions are usually scattered. It is caused by cancerous growth of plasma cells present in the bone marrow. They replace healthy cells, causing gradual bone destruction. This process is called tumour osteolysis. Bone losses and fractures occur as a consequence. The protein produced by the myeloma cells enters the bloodstream, leading to renal failure. Because the bone marrow is also responsible for the production of red blood cells and immune cells, patients often have anaemia and are susceptible to infection. Myeloma is a chronic disease. Its first stage is asymptomatic, and subsequent ones may lead to the development of chronic renal failure, pathological bone fractures, and bone marrow failure. It is therefore important to quickly diagnose a disease so that appropriate therapy can be started as early as possible.
It is important in diagnosis to precisely determine the severity of the disease. This is achieved by diagnostic tests including bone marrow biopsies, measurement of serum monoclonal protein (paraprotein) and/or urine (Bence-Jones protein) and bone testing. In the latter case, radiological examination and densitometry are most often performed. Radiological test allows to assess the presence, progression and localisation of bone damage [1]. Characteristic defects (so-called cheese holes) in the bones are called osteolytic foci. These are the places where the myeloma destroyed the bone (Fig. 1). Bone loss is confirmed by densitometry. Of course, other imaging methods such as magnetic resonance imaging (MRI), computed tomography (CT) or positron-emission tomography (PET) are also used. These methods are more accurate but they are used more often to precise imaging all damaged bone fragments, when the disease has already been diagnosed with laboratory tests [2].
Currently, radiological examination is a standard diagnostic method (the so-called “gold standard”) for bone damage. It detects evidence of osteolytic changes, osteopenia, lytic lesions and fractures. There is a relationship between the severity of the disease (number of lytic lesions in the patient) and the amount of tumour mass during diagnosis [3]. The radiological test is cost-effective, widely available, allows for the visualisation of large areas of the skeleton and can identify long bones potentially at risk for fractures in osteolysis sites. The main disadvantage of radiological examination is the low sensitivity of this method, allowing only lytic lesions to be visualised in the case of losing at least 30% of the spongy bone [4]. In addition, this method provides an inadequate assessment of generalised osteopenia [5] and is characterised by low specificity, particularly within the spine [6]. Radiological examination may reveal non-specific abnormalities that require further identification by other means (e.g. CT) and some parts of the skeletal system are difficult to evaluate precisely (e.g. breastbone, ribs and shoulders).
In case of MM, radiology is primarily used in the early stages of the disease, in the detection and characterisation of complications, and in the assessment of the patient's response to treatment. It has been found that radiological examination of the skeletal system permits a more accurate assessment of bone disease progression in patients with myeloma than limited MR imaging [7]. Unfortunately, as stated earlier, the main disadvantage of the method is its low sensitivity and specificity. On the other hand, it is important to detect the disease as early as possible because implementing appropriate early-stage treatment procedures increases the chances of effective treatment. Therefore, it would be beneficial to increase the accuracy of the diagnosis based on the radiological examination. Standard diagnosis of myeloma is based on the results of laboratory tests. Detailed guidelines for diagnostic criteria and scope of examinations are included in the report of the International Myeloma Workshop Consensus Panel 3 [8]. This report has been updated by the International Myeloma Working Group in 2014. Many works present the results of own research relating to the above criteria [9], [10], [11]. In the light of current research, the sensitivity of disease detection can be increased by using advanced cross-sectional imaging techniques, such as MR and FDG PET/CT [12], [13], [14], [15], [16], [17]. Whole-body scintigraphy is also used, which allows for more accurate distinction between mild osteolytic lesions and disseminated MM metastases [18]. Unfortunately, the diagnosis of the disease as a result of laboratory testing takes place very often in its advanced stage. Therefore, widely available tools are needed to support the process of initial diagnosis and increase the probability of early detection of the disease. It seems that such a task could be fulfilled by a method using appropriate image processing algorithms based on computer radiography.
There are relatively few publications presenting the results of the works concerning computer methods for biomedical image processing and analysing aimed at diagnosing MM. Some methods use the bone marrow for this purpose. This approach is described in Ref. [19] where the SVM classifier was used to diagnose the disease based on microscopic bone marrow images. A method for detecting myeloma with the probabilistic model of bone marrow is presented in Ref. [20]. In turn, a probabilistic, spatially-dependent density model of normal tissue was described in Ref. [21]. In the above two cases, the analysis was based on femur CT images. In Ref. [22], the results obtained by multimodal MRI image analysis are presented. In this work, classification by the method of random forests for the detection of myeloma in lumbar vertebrae was used. Hering et al. [23] applied generalised multiple instance learning to automatic detection of lesions caused by MM in femurs. They used CT images. The aim of the research presented in Ref. [24] was to build a classifier for detecting changes caused by MM. Patients were previously diagnosed based on CT images. The authors built a classifier based on multimodal MRI images of the pelvic bone. The authors are unaware of studies in which the results of computed radiography (CR) image texture analysis in diagnosing MM have been used.
A review of the literature shows that research is carried out on the methods for automatic detection of myeloma, which use more accurate imaging methods than CR. They use, e.g. MRI or CT. However, the main drawback of these solutions is the fact that they are more expensive and less available than CR. In addition, they are more often used to accurately assess the damage of the selected bone structure fragments, after having previously diagnosed the disease with laboratory tests. On the other hand, radiology is still used in the early stages of the disease. Therefore, in our opinion, it is worth undertaking a research effort, which is an attempt to improve the accuracy of the radiological examination through the use of appropriate algorithms of computer image analysis. The advantages of using CR images are that this approach is cost-effective and noninvasive. In addition, the radiological examination is widely available and is a standard diagnostic approach to assess the presence, progression and localisation of bone damages. In the paper, we presented a new approach that is the use of textural features of the humerus CR image in detecting of the bone damage caused by the aforementioned disease.
The article consists of 5 sections. Section 2 describes the material and methods used, including image pre-processing, data standardisation and feature selection. In Section 3, the results are presented. The way of classifier building and testing is discussed in Section 4. And finally, conclusions are presented in Section 4.
Section snippets
Material
In the study, CR images belonging to 26 patients were used. Thirteen persons were healthy and thirteen patients were diagnosed with myeloma. The ground truth of whether a given CR image belongs to the healthy or sick person was determined by a haematologist (an expert in the field) based on clinical tests. The full set of images taken during the radiological examination included images of the skull, thorax, spine (cervical, thoracic, lumbar), pelvic bone, humerus and femur. All the images were
Results
The full set of 279 features was subjected to statistical analysis. On this basis, 152 features significant at the level α = 0.01 were found. Then, they were sorted according to the correlation value with the dependent variable (sick or healthy). Based on the ranking, the first 23 features of the highest correlation were chosen for further selection by Hellwig's method (Table 1).
The maximum number of features in the analysed combinations was set to 10. As a result of the implementation of
Discussion
It is hard to compare own results with the results of other research, because (as noted in Section 1) the authors are not aware of the works where CR textures were used in the process of automatic diagnosis of MM. Also, there are a small number of publications presenting the results of the works where automatic diagnosis of the disease is based on analysis of images other than CR images. In Ref. [19], the classification results of bone marrow microscopic images using the SVM classifier are
Conclusions
The results of the study confirm the thesis that humerus CR images can be used in the detection of bone damages caused by multiple myeloma. The classification accuracy obtained for individual quality indexes was as follows: ACC = 93%, TPR = 92%, TNR = 96%, PPV = 96% and NPV = 93%. Achieving such accuracy was possible thanks to the Hellwig's method, which confirmed the high efficiency of the selection of the analysed textural features in question. Two feature combinations have the best ability to
Authors’ contributions
Zbigniew Omiotek and Małgorzata Szatkowska wrote the manuscript; Zbigniew Omiotek prepared the original draft, also involved in methodology and investigation of the study. Olga Stepanchenko made the visualisation and software analysis. Waldemar Wójcik conceptualised and supervised the work. Data curation was performed by Wojciech Legieć. Małgorzata Szatkowska reviewed and edited the manuscript.
References (43)
- et al.
Consensus recommendations for standard investigative workup: report of the International Myeloma Workshop Consensus Panel 3
Blood
(2011) - et al.
Myeloma bone and extra-medullary disease: role of PET/CT and other whole-body imaging techniques
Crit Rev Oncol Hematol
(2016) - et al.
MR image texture analysis – an approach to tissue characterization
Magnet Reson Imaging
(1993) Fractal analysis of the grey and binary images in diagnosis of Hashimoto's thyroiditis
Biocybern Biomed Eng
(2017)- et al.
Choice of the smoothing parameter and efficiency of the k-nearest neighbor classification
Comput Math Appl
(1986) - et al.
Data mining techniques and applications –a decade review from 2000 to 2011
Expert Syst Appl
(2012) - et al.
Bundling classifiers by bagging trees
Comput Stat Data Anal
(2005) - et al.
Abnormal radiological features in a multiple myeloma patient: a case report and radiological review of myelomas
Dentomaxillofac Radiol
(2011) - et al.
Multiple myeloma: a review of imaging features and radiological techniques
Bone Marrow Res
(2011) - et al.
A clinical staging system for multiple myeloma. Correlation of measured myeloma cell mass with presenting clinical features, response to treatment, and survival
Cancer
(1975)