Although classification of astrocytic tumors is standardized by the WHO grading system, which is mainly based on microscopy-derived, histomorphological features, there is great interobserver variability. The main causes are thought to be the complexity of morphological details varying from tumor to tumor and from patient to patient, variations in the technical histopathological procedures like staining protocols, and finally the individual experience of the diagnosing pathologist. Thus, to raise astrocytoma grading to a more objective standard, this paper proposes a methodology based on atomic force microscopy (AFM) derived images made from histopathological samples in combination with data mining techniques. By comparing AFM images with corresponding light microscopy images of the same area, the progressive formation of cavities due to cell necrosis was identified as a typical morphological marker for a computer-assisted analysis. Using genetic programming as a tool for feature analysis, a best model was created that achieved 94.74% classification accuracy in distinguishing grade II tumors from grade IV ones. While utilizing modern image analysis techniques, AFM may become an important tool in astrocytic tumor diagnosis. By this way patients suffering from grade II tumors are identified unambiguously, having a less risk for malignant transformation. They would benefit from early adjuvant therapies.
Following the classification of the World Health Organization (WHO) astrocytic tumors (gliomas) are divided into four grades, which are typically assigned on the microscopic appearance of the tumor [
Atomic force microscopy (AFM) has become a widely used technique for characterizing biological samples at nanometer resolution. In a lot of studies [
A common method to decide grading of astrocytoma is the examination of H&E-stained histological brain slices with a light microscope [
Accurate classification of brain tumor grading is very important in the diagnosis because it defines prognosis and treatment decision for the patient. Dependant on the used standard imaging technique many refined methods were developed to increase grading accuracy [
In this study, we present an objective method, which is well suited to enhance the accuracy in the determination of specific tumor features. The presented method differs from other ones by the combination of two key elements: (a) high-resolution microscopy where we will show that AFM imaging on histological unstained brain samples is able to deduce relevant morphological information, which can be used for grading astrocytoma; (b) image analysis where we will demonstrate that the application of special data mining algorithms based on Minkowski functionals enables an objective, automatic identification of histomorphological features also in such a complex task like astrocytoma grading. This automatic approach enables improved classification accuracy in the future and could offer new diagnostic elements for an objectivized morphological tumor categorization.
The clinical material comprised brain tumor samples from 14 patients that were made available from the pathological institute of the Nerve Clinic Wagner-Jauregg in Linz (Austria). The samples were classified as low (astrocytoma grade II,
The samples were prepared according to standard pathology protocol [
The AFM measurements were performed on an Agilent 5400 AFM/SPM (Agilent Technologies Inc., Santa Clara, CA, USA), equipped with a large multipurpose scanner and a digital camera (Navitar Inc., Rochester, NY, USA). All images were acquired in air at room temperature in contact mode using commercial non-conductive silicon nitride cantilevers (Bruker Corporation, Camarillo, CA, USA) with a spring constant between 0.005–0.06 N/m. The images were taken at 512 × 512 pixels quality at a scanning rate of 1.0 lines/second. All images were recorded with PicoView 1.4.2 (Agilent Technologies Inc., Chandler AZ, USA) and further processed with Pico Image Basic 5.0.4.5170 (Agilent Technologies Inc., Chandler, AZ, USA). Altogether the analysis comprised images of 113 samples (54 astrocytoma grade II and 59 glioblastoma multiforme grade IV).
Light microscopy was performed on a Nikon eclipse ME 600 (Nikon Instruments Austria, Vienna). The magnifications were 50, 100, and 500.
Figure
The main processing steps in image analysis.
Histogram equalization for a typical astrocytoma grade II AFM image. (a) Original AFM image composed of 512 × 512 pixels. The corresponding scan size was 100 × 100
Histogram equalization for a typical glioblastoma multiforme grade IV AFM image. (a) Original AFM image composed of 512 × 512 pixels. The corresponding scan size was 100 × 100
We finally used Minkowski functionals (or to be more precise “Minkowski measures”)—in particular the Euler-characteristic—as a feature descriptor to characterize global geometric structures related to the topology of the AFM images. The Euler-characteristic is defined as the total number of objects in an image minus the number of holes in those objects. Figure
Illustration of the Euler-characteristic. The Euler-characteristic is defined as the total number of objects in the image minus the number of holes in those objects. Exemplarily, the Euler-characteristic of the given binary image is equal to −2.
In two dimensions, the Minkowski functionals are related to more familiar quantities like the covered image area, the boundary (or contour) length between homogeneous domains, and the Euler-characteristic (i.e., the number difference of connected domains and holes). Here, we focused on Minkowski functionals to characterize the morphology of image domains that result from thresholding AFM height maps at different height levels (i.e., binarization of the AFM image at different gray levels to transform the AFM height map to a stack of level sets).
Figure
Original AFM image of an astrocytoma grade II, and 5 image examples of binarization corresponding to the height threshold levels 32, 64, 128, 192, and 224.
Original AFM image of a glioblastoma multiforme grade IV and 5 image examples of binarization corresponding to the height threshold levels 32, 64, 128, 192, and 224.
To prove AFM as a morphological tool in pathology images, brain tumor specimens with light microscopy stained by routine H&E were compared with our AFM results. Typical glioma features like pleomorphic cells or the pseudoglomerulus endothelial proliferation could be recognized very clear in AFM images. Figure
Typical images of brain tumor samples. (a) Light microscopy image of a stained astrocytoma grade II sample at 100x magnification. Healthy microglia (small black arrows), lymphocytes, and tumor cells (large black arrows) can be identified. (b) Light microscopy image of a glioblastoma multiforme grade IV at 100x magnification. The neuropil is highly porous. Most of the polymorphous tumor cells are surrounded by large cavities (black asterisks). Two isolated blood vessels are apparent by erythrocytes and atypical endothelial cells (large red arrows). (c) Corresponding AFM image at 98 × 98
By comparing images of low-grade and high-grade tumors the gradual loss in fine regular anatomy of the neuropil appeared as a noticeable new characteristic tumor feature, because it occurred in close accordance with the tumor type and grading. This gradual loss, which is consistent with a tumor associated loss of functional organization, is accompanied by an increase in neuropil-free areas, which appear dark in the AFM images. Thus, the formation of dark areas was taken as the key feature for the further processing in determining the grading of brain tumors.
Figure
The resulting mean value (solid line) of the Minkowski functional Euler-characteristic and the 1
Another topological descriptor is the Minkowski functional contour length, which is also plotted for both tumor types (astrocytoma: red curves, and glioblastoma multiforme green curves) in Figure
The Minkowski functional contour length and the corresponding 1
An accurate classification of brain tumors is of utmost importance, because it is the basis for an optimal therapy. The search for new grading markers is necessary to improve personalized therapies in a devastating disease like high-grade brain tumors. The WHO has published a classification scheme which is used worldwide for neuropathological typing and grading of brain tumors. The scheme is mainly based on histomorphological features [
In our study, we used AFM on routine brain tumor samples. Tumor diagnosis and tumor grading were performed by experienced neuropathologists. Artifact-free specimens were selected, showing characteristic tumor features, and the adequate paraffin block was chosen for routine microtome slices 5
The total dataset comprises 113 samples, containing 54 samples of astrocytoma (Grade II) and 59 samples of glioblastoma (Grade IV). Thus, the dataset is nearly balanced which is important for validating the significance of obtained classification accuracies.
For a data-modeling process, creation and selection of appropriate features are essential to the reachable model accuracy. In particular when using image processing for classification, these steps majorly influence the achievable classification results because extracting features from images concern extracting the highest possible amount of information. Here, the given Euler-characteristics provide some very useful data that shall be applied for classification. When using Euler-characteristics directly for classification, for each sample 256 features would need to be considered (i.e., the characteristic’s values at 256 gray levels), which gives an unfavorable sample size to dimensionality ratio. However, even if [
When aiming at creating generalizable models out of data, the distinction of proper training and test sets is fundamental. Since (as in most biologic applications where measurement costs are high and samples are difficult to obtain) the sample size is very limited, cross-validation is applied for classification. Cross-validation is the method of dividing the available data into
Genetic programming (GP) is an evolutionary algorithm-based method for symbolic classification. It produces interpretable models that allow the assessment of the impact of each single feature in the Minkowski functional curves by this way optimizing the resulting classification accuracy [
Parameter settings for best GP result.
Parameter | Tested values |
---|---|
Maximum generations | 70, 100 |
Mutation probability [%] | 10, 15 |
Population size | 70 |
Selector Mutator | Tournament sector |
Maximum expression tree depth | 8, 10 |
Maximum expression tree length | 25, 50, 80 |
Symbolic expression tree grammar | Logical operators (the corresponding expression tree symbol set implemented in Heuristic-Lab contains too numerous functions to be mentioned here. Generally, it contains all usually handled logic operators). |
GP’s best model resulted in 0.93% better classification accuracy by reducing the dimension
A closer view at the Minkowski functional Euler-characteristic revealed a noteworthy, discriminative detail. Between the gray levels 80 and 105, there is a region where the confidence intervals of both mean curves do not intersect. This fact was additionally considered algorithmically using GP and the analysis of this single feature resulted in a prediction accuracy of 89.38%. Obviously, this feature alone has a very high-discriminative capability. Maybe it is a new additional key feature to the conventional grading procedure to separate tumors that are difficult to distinguish, for example, grade II tumors from grade III ones or grade III tumors from grade IV ones. It is also conceivable that this special feature will help in understanding unusual courses of illness. In our analysis, about 10% of the data did not fit the above mentioned prediction accuracy. These data has to be analyzed in the future in a medical orientated paper. Despite the high accuracy of our grading tool the 10% have to be correlated with special tumor features, the tumor region, and finally with the patient history and the tumor outcome. Tumor discovery is mainly based on medical imaging techniques (MRI, CT, or PET), and tumor diagnosis is done by histopathological examination on biopsied or resected tumor tissues. To achieve highest accuracy in classification data mining techniques have been successfully performed in some special cases, especially in MRI and light microscopy images. Reference [
Depending on the used imaging method and the performed data mining technique, classification accuracies between 90.3% and 97.3% were achieved, only taking values concerning the differentiation between different tumor types into account. The proposed AFM-based classification method using GP as classifier achieves 94.74%, which is absolutely comparable with the literature data. AFM has one additional advantage above all concerning the analysis of H&E-stained images. The quality of the images does not depend on the staining protocol, because objects are not viewed, but rather scanned. Thus, variations in color are not detected, which makes AFM insensitive with respect to varying staining conditions and artifacts.
Imaging histological brain samples using AFM, we were able to discover a new diagnostic feature, the neuropil density, which is a perfect structure for modern computer-based imaging analysis. In combination with data mining techniques this characteristic feature can be used to raise astrocytoma grading to a more objective standard to improve classification accuracy in combination with conventional pathological tumor diagnosis and grading procedure. Additional morphological criteria are of great value to subdivide known tumor entities and to find new grading criteria for therapy decisions. AFM is easily performed on routine pathological samples without additional processing steps. Therefore, the method is easily integrated in the daily routine; AFM could be performed after deparaffination and before the sample staining procedure without diagnostic relay. A further advantage is that AFM is able to analyze sections of old paraffin tumor blocks stored in many pathological labs. Therefore, the method is also useful for retrospective studies on well-defined tumor collections and clinical data.
Brain tumor classification based on AFM images by using GP has not been done so far. It is a new methodology bringing high-resolution microscopy closer to the clinical practice. It is able to achieve a classification accuracy, which matches or outperforms most of the proposed techniques described in the literature. Additionally, the potential of imaging in the submicron regime enables the characterization of ultrastructures as new diagnostic details within samples in routine pathology that are not visible using conventional medical imaging techniques or light microscopy in tumor diagnostics. AFM is easily implementable in the diagnostic process. Together with data mining techniques, AFM could serve as a powerful new tool in pathological diagnosis and in objectifying morphological features for tumor diagnosis and grading.