Differential diagnosis of focal changes in the spine using standard and radiomic analysis

Full Text

BACKGROUND

Bone metastases have been reported to occur in approximately 40% of patients with cancer, with varying incidences depending on the primary tumor type and location [1]. Contemporary medical imaging methods, such as multislice computed tomography, magnetic resonance imaging (MRI), and positron emission tomography, have high diagnostic efficiency. However, the variability in the course of bone metastases determines the advantages and disadvantages inherent in each method [2, 3]. Modern image analysis algorithms aim to improve the interpretation of the obtained research results. Patients’ conditions can be described in various ways and with varying degrees of detail based on MRI images using different methods, such as pixel-by-pixel analysis of images for various pathologies and modalities [4, 5]. Here, several methods can be selected, which are mainly reduced to certain statistics, namely, brightness statistics, histograms, and distribution of pixel groups (i.e., texture). Because of such large variability in values, despite attempts to systematize and unify the approach, at this stage, many different methods of analysis often provide contradictory or doubtful results [6]. Gradually, with the improvement in the methods of recognizing an object (disease), an increasing number of signs are included, leading to direct recognition of objects by a person or a neural network [7, 8].

Therefore, this study aimed to improve the accuracy of diagnostic radiographic findings of focal changes in the spine using an additional image evaluation by computer analysis algorithms.

MATERIALS AND METHODS

Study design

This was a single-center retrospective study. A total of 30 patients were included in this study. MRI data of the patients with focal changes in the spine were independently analyzed by two radiologists with >10 years of experience. If the conclusions of the radiologists coincided, the results of the images were divided into three groups (hemangiomas, metastasis, and normal bone marrow), after which they were subjected to radiomic analysis with the identification of features characteristic of each group.

Terms and conditions

MRI images were retrospectively analyzed on the workstation “PO Vidar” (Russia). Images were analyzed using the specialized software “Radiomica Applata” (Germany). MRI examinations were performed using standard protocols approved by the Russian Scientific Center of Roentgenology and Radiology (Moscow).

Methods for assessing target indicators

Research methods

The diagnostic algorithm used MRI from devices from various manufacturers (Aera, Siemens, Germany; Atlas, Canon, Japan) with a magnetic field strength of 1.5 T. All patients with cancer underwent bone scintigraphy within the examination protocol. All studies were independently evaluated by two radiologists with >10 years of experience to increase the significance.

Analysis methods

The key MRI sequence for radiomic analysis was T₂-weighted images, and T₁-weighted images and short tau inversion recovery (STIR) were used as auxiliary ones. After image normalization, a pixel-by-pixel analysis was performed with an assessment of the shape and an analysis of the histograms of the image brightness distribution [9]. When using these algorithms, sufficient homogeneity of the object is considered an important parameter. However, this requirement is strong but not always appropriate for many applications. Thus, as it should be in diagnostics, especially in differential diagnostics, additional information about the object should be included while limiting applicability [10].

On the desktop of “Radiomica Applata” (Germany, a proprietary design of IT researchers), the object under study was isolated in the histogram mode, which displays the frequency of occurrence of brightness graphically in the grayscale and enables setting the upper and lower threshold values. For example, when analyzing a hemangioma, if the corresponding brightness window is taken on the histogram, poorly distinguishable objects of a homogeneous structure can be selected (Fig. 1).

 

Fig. 1. T₁ tse sag-hemangioma, highlighted with a window on the brightness histogram, which is shown on the right side of the figure.

 

As the next extension of the object description, angiogenesis of breast cancer (BC) metastases was used, which has an additional property, namely, a chaotic pathological growth of the vascular network, compared with normal tissue. This determines the image complexity, where the object appears to be more chaotic, i.e., less complex than another, if the sum of boundaries of the parts of its components is less than that of the other under the same measurement conditions. The sum of the boundaries is calculated by the action of the gradient operator on the image under study, that is, a function of pairs of pixels (Arzela operator).

For shape recognition and the use of other modalities, neural networks with the architecture of a multilayer perceptron were used with an input vector of 100 × 100 [11]. The examples described do not require an analysis of the separability of the set of training images (Fig. 2).

 

Fig. 2. Multilayer perceptron for recognition in T₂.

Note (here and in Fig. 3). hema — hemangioma, mts — metastasis, norm — normal tissue.

 

In other cases, a self-learning Kohonen network was used for this purpose. The following groups of main images were taken to train the networks:

• Hemangioma.
• Normal tissue.
• Metastases.

Spondylocystitis and areas of fatty degeneration in the vertebrae or the so-called fat deposits, namely, local focal changes with a high-fat content in the presence of intact bone marrow, were used as additional groups.

Ethical considerations

Ethical approval was waived due to the retrospective nature of the study.

RESULTS

Participants

The results of the examination of 30 patients with focal changes in the vertebral bodies of the thoracic and lumbar spine were included in this study. Of the 30 patients, 15 had morphologically verified BC with metastatic focal changes in the bones, and 15 did not have a morphologically confirmed diagnosis of a malignant neoplasm, and they were examined due to nonspecific complaints of “pain or discomfort in the spine.” The most diagnostically indicative foci were used in the study. In this category of patients, hemangiomas of the vertebral bodies (n=7) and areas of subchondral edema in the vertebral bodies (n=5, 3 of them had spondylodiscitis) were determined. Furthermore, for comparison, areas of local delimited fatty degeneration of the bone marrow were analyzed (n=5), and areas of unchanged vertebrae were measured in five patients. A total of 40 measurements were made in this study.

Primary results

The values of the ratios of light, medium, and dark intensities were obtained when analyzing the intensities in the studied objects. Dark sites were defined as microfoci or the so-called calderas. The ratio of the average brightness of an object (hemangioma, metastasis, or an area of conditionally unaltered bone marrow) to the brightness of the surrounding bone tissue of the vertebral body in grayscale was chosen as a “histogram” assessment. Table 1 shows the radiomic analysis results of hemangiomas. Table 2 shows the analysis of BC metastatic foci with the closest possible clinical characteristics. In the table, microfoci with brightness significantly higher than the average in the region of interest were marked in yellow, those with signal isointense (ISO) to the environment were marked in red, and areas with reduced intensity compared with the average were marked in green.

 

Table 1. Results of radiomic analysis of hemangiomas depending on the operators used in T₂WI sag mode

Description of RoI	Arzela RoI	RoI caldera	Arzela data image of the surroundings	RoI/ surroundings in gray
Hemangioma-1, Т2-sagittal				1.27
Hemangioma-1, Т2-sagittal, subtotal				1.16
Hemangioma-2, Т2-sagittal				1.16
Hemangioma, Т2-sagittal				1.11
Fatty degeneration, Т2-sagittal				1.07
Hemangioma-2, Т2-sagittal				1.17

 

The hemangioma manifested itself in increased brightness compared with normal tissue, with a ratio of 1.11:1.27 in the cases discussed, increased structuredness (multiple ISO microfoci), and increased complexity (Arzela) to normal tissue. Furthermore, cases of fatty degeneration were reported, the main difference of which, compared with hemangioma, was the environment. Our parameters (intensity statistics, pattern complexity, and the presence of microfoci, which might not be present) did not reflect the environment and imaging mode (e.g., fat suppression) and were thus only necessary but insufficient.

BC metastases appeared to be reduced in brightness compared with normal tissue on a T-weighted image with a ratio of approximately 0.9 (Table 2). Additionally, a certain increase in image complexity was noted. The caldera was distinctly elevated and varied from ISO to darkened, which is expressed as a percentage of the total image area. The study results showed that the main differential indicators of hemangioma, conditionally normal areas of the bone marrow, and metastatic foci were established. The Arzela image complexity operator was approximately 0.07 for hemangiomas, approximately 0.05 for metastases, and approximately 0.04 for vertebrae. The histogram brightness operator was approximately 1.12 for hemangiomas and approximately 0.94 for metastases. The difference between the indicators was about 20%–25% between hemangiomas and bone marrow and 35% between metastases and bone marrow. Therefore, these indicators can be effectively used together with other markers.

 

Table 2. Results of T2WI radiomic analysis of breast cancer (РМЖ) metastases

BC metastasis 1		0.915
BC metastasis 2		0.965
BC metastasis 3		0.945
BC metastasis 4		0.88
BC metastasis		0.98

 

Figure 3 shows the result of object recognition from Fig. 1. The Arzela value was superior to that of the “normal” adjacent tissue (leftmost image), which is typical for a hemangioma. Additionally, the ratio of average intensity to “normal” tissue, equal to 1.15, was characteristic of hemangioma. The conditional probabilities of identifying the object were indicated, from which the hemangioma was recognized, with a conditional probability of 0.9 in shape.

 

Fig. 3. The results of the analysis of hemangioma by the Artzel drawing complexity operator.

 

Furthermore, as a control measure, the same parameters in the STIR MRI sequence were analyzed, which is considered an important diagnostic component in the differential diagnosis of focal changes in the spine. Here, the hemangioma also manifested itself to be brighter than normal tissue, with a ratio between 1.02 and 1.24 in the cases under consideration. The increased structuredness and increased complexity of the Arzela operator in relation to normal tissue were important established criteria.

The image analysis operators used did not reveal significant differences in the differential signs of the presented edema and vertebral fat deposit cases. This is probably due to parallel processes, where areas of fatty degeneration with a reduced signal in STIR and uneven zones of subcortical osteosclerosis are determined against the subcortical peritrabecular edema (Fig. 4).

 

Fig. 4. MRI of the lumbar spine in the sagittal projection. T₂WI, degenerative-dystrophic changes in a patient without an oncological history. a) Zones of heterogeneous subcortical edema are determined in the bodies of the vertebrae L_IV, L_V against the background of a herniated disc. b) In the STIR mode, the severity of edema appears to be reduced in relation to T₂WI due to areas of fatty degeneration and subcortical sclerosis.

 

DISCUSSION

Computer analysis methods of images, such as radiomics, texture analysis, and computer aid systems, have great prospects. Some solutions are already being used in practice, such as in screening for focal changes in the lungs, predicting the results of treatment of brain tumors, and direct evaluation of treatment efficiency of BC metastases [12–14]. Although this study demonstrated a rather high efficiency in distinguishing between benign and malignant focal changes in the spine, this can be due to the fact that the clinical cases most typical from a radiographic point of view were analyzed in this study. In practice, several parallel chronic processes pose the greatest difficulty for radiologists in evaluating bones when changes can overlap. Improving the diagnostic efficiency of the method consists of cross-analysis of various sequences and operators and search for new radiomic image markers, which requires further research and a larger number of cases.

Some measurement errors may depend on the difference in image filters implemented on devices from different manufacturers. Thus, more cases and statistical analyses of the results obtained are needed to increase the significance of the results obtained.

CONCLUSION

The differential diagnostic criteria demonstrated in this study, including data obtained using neural networks, are necessary but insufficient from a mathematical point of view and are advisory in nature. Medical specialists with experience are still at the core of the decision-making system.

ДОПОЛНИТЕЛЬНО / ADDITIONAL INFO

Вклад авторов. Все авторы подтверждают соответствие своего авторства международным критериям ICMJE (все авторы внесли существенный вклад в разработку концепции, проведение исследования и подготовку статьи, прочли и одобрили финальную версию перед публикацией).

Author’s contribution. Thereby, all authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work.

Источник финансирования. Не указан.

Funding source. Not specified.

Конфликт интересов. Авторы декларируют отсутствие явных и потенциальных конфликтов интересов, связанных с публикацией настоящей статьи.

Competing interests. The authors declare that they have no competing interests.

About the authors

Nikolay I. Sergeev

Russian Scientific Center of Roentgenoradiology

Author for correspondence.
Email: sergeevnickolay@yandex.ru
ORCID iD: 0000-0003-4147-1928
SPIN-code: 2408-6502
Scopus Author ID: 57201131922
http://www.rncrr.ru

MD, Dr. Sci. (Med.), leading researcher

Russian Federation, 86 Profsoyuznaya Str., 117485 Moscow

Petr M. Kotlyarov

Russian Scientific Center of Roentgenoradiology

Email: marnad@list.ru
ORCID iD: 0000-0003-1940-9175
SPIN-code: 1781-2199

MD, Dr. Sci. (Med.), Prof.

Russian Federation, 86 Profsoyuznaya Str., 117485 Moscow

Vladimir A. Solodkiy

Russian Scientific Center of Roentgenoradiology

Email: director@rncrr.ru
ORCID iD: 0000-0002-1641-6452
SPIN-code: 9556-6556

MD, Dr. Sci. (Med.), Prof.

Russian Federation, 86 Profsoyuznaya Str., 117485 Moscow

References

Supplementary files