Introduction

Amide proton transfer weighted (APTw) imaging represents a novel contrast media free molecular MR imaging technique that has recently shown promise in characterizing and differentiating brain neoplasms as well as malignancies in other body regions1,2,3,4,5,6,7. The APTw signal originates from amide protons in endogeneous proteins and peptides in the parenchyma. In tumor tissue, the content of mobile proteins and peptides is increased thus resulting in increased APTw signal intensity values8,9.

With the exception of one study10, these previous investigations utilized standard histogram analyses techniques at the most to analyse the APTw signal of the tissue at hand thus only scratching the surface of the information that can potentially be extracted from radiological images1,2,3.

With recent advances in the field of machine learning (ML), radiomics techniques allowing for the extraction of high-dimensional mineable data from medical images have been developed and introduced to medical imaging thus enabling in-depth tissue classification and characterization11,12,13,14,15.

In this proof-of-concept study we aimed at assessing the potential of radiomics and ML for APTw imaging. To this extent, we used radiomics on APTw images to differentiate WHO Grade 2, 3 and 4 gliomas from brain metastases.

Materials and methods

This study received institutional review board approval (Cantonal Ethical Committee Zürich, Switzerland) and was performed in accordance with all guidelines and regulations defined by the institutional review board. All subjects gave written informed general consent.

Subjects

In this retrospective study we included 48 patients (mean age: 61 years, range: 37–83 years) diagnosed with either low- or high grade gliomas (10 patients with WHO Grade 2 tumors, 1 patient with WHO Grade 3 tumor and 10 patients with WHO Grade 4 tumors (= glioblastoma)) or brain metastases (MET; 27 patients, with 15 lung, 1 breast, 9 melanoma and 2 kidney as primary sites of origin). Initially 64 consecutive patient studies acquired between august 2018 and april 2020 at a single tertiary institution were reviewed. Then the following exclusion criteria were applied: < 18 years, lack of histological analysis of neoplasm (6 patients), unavailable MRI data (1 patient), treatment prior to MRI, lesions with a diameter of under 10 mm16 (3 patients), failure to compute APTw signal intensities during image reconstruction (6 patients). In the case of multiple metastases all lesions were used for analysis (if they fulfilled size requirements) and were pooled17 thus resulting in one data point per patient. A flowchart can be found in the supplementary material.

MR imaging

Subjects were examined on a clinical 3 T scanner (Achieva, Philips Healthcare, Best, the Netherlands) with an eight channel receive-only head coil array. The following sequences were acquired: precontrast transverse diffusion-weighted imaging (DWI), 3D double inversion recovery (DIR), 3D T1-weighted (T1w) turbo field echo (TFE) and postcontrast 3D fluid attenuated inversion recovery (FLAIR), 3D T1 black blood turbo spin echo (TSE) or 3D T1w m-Dixon TFE, APTw and 2D T2w TSE.

Table 1 shows the sequence parameters for the APTw sequence.

Table 1 APTw MRI sequence parameters.
Full size table

We used a clinically approved APTw sequence18,19,20 acquired in transverse oblique orientation parallel to the intercommissural line. 16 slices with a slice thickness of 3.85 mm were acquired. The first slice was centered at the inferior border connecting the rostrum and the splenium of the corpus callosum19,20. Details on how APTw imaging contrast was generated can be found in the supplementary material.

Radiomics feature extraction and image analysis

APTw and postcontrast T1w or FLAIR Digital Imaging and Communications in Medicine (DICOM) files were loaded into the open-source software platform 3D Slicer (v. 4.10.2) and were aligned geometrically. Subsequently, two readers (TS and ES each with 3 years of experience) manually segmented the neoplasms on overlayed images according to contrast enhancement and solid parts of neoplasms21. Segmentation was performed on all axial slices for 3D segmentation of either T1w postcontrast or FLAIR images superficially overlaid onto the APTw images (Fig. 1). Specifically, readers outlined neoplasms on pseudo-images that were generated by accurately overlaying APTw with the now geometrical identical structural images. These pseudo-images contain the APTw imaging information but visually reflect T1w postcontrast or FLAIR images. Intensity discretization was performed to a bin width of 25. Gray level cooccurrence matrix (GLCM) features were computed at 4 inter-pixel distances.

Figure 1
figure 1

Representative image examples of 4 patients who presented with different neoplasms. FLAIR or T1w postcontrast images are provided in the top row, while the corresponding APTw images are shown in the bottom row.

Full size image

Then 110 radiomics features were extracted with the built-in pyRadiomics package implemented into 3D Slicer22. Most features are in accordance with those described in the Imaging Biomarker Standardization Initiative (IBSI)23. Radiomics features corresponded to seven different matrices/feature classes: First-order statistics/histogram matrix, shape-based features, gray-level cooccurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighbouring gray tone difference matrix (NGTDM), and gray-level dependence matrix (GLDM). A detailed overview and description of radiomics features can be found elsewhere22,24.

Radiomics features—dimension reduction

Dimension reduction was performed in two steps. First, radiomics features of both readers were compared by means of intraclass correlation coefficients (ICC). ICC values of greater than 0.8 were interpreted as excellent agreement11,25. Radiomics features with ICC values below this threshold were discarded from further analysis, as shown previously26. In a second step, a classifier attribute evaluation filter (CfsSubsetEval) of the open source software package Weka (WEKA, version 3.8.3, University of Waikato, Hamilton, New Zealand) were applied on the training data to evaluate the worth of an attribute. This method measures the significance of attributes on the basis of predictive ability of attributes and its degree of redundancy. The subsets which are having less intercorrelation but are highly correlated to the target class are selected for further analyses.

The remaining radiomics features were then used to train ML classifiers. Combinations of the weighted radiomics features were used then to distinguish metastases from glial primary brain tumours.

For the consecutive subanalyses of distinguishing glioblastomas from other glial brain tumors and metastases, we performed a prior principal component analysis to cover approximately 95% of variance in the original dataset.

Machine learning

For ML analysis, open-source software (WEKA, version 3.8.3, University of Waikato, Hamilton, New Zealand) was used. For prediction of histopathology, a commonly-used ML algorithm implemented in the open source WEKA package was tested with handpicked hyperparameters: Multilayer perceptron, which uses backpropagation to learn a multi-layer perceptron to classify instances with a learning rate of 0.3 and a momentum of 0.2. For further subanalyses we used a random forest classifier. All results were tenfold cross validated to overcome overfitting.

Results

Dimension reduction

After dimension reduction regarding reproducibility and attribute evaluation, eight out of 110 radiomics features remained for further analysis (Table 2).

Table 2 Radiomics features after dimension reduction.
Full size table

Machine learning

The Multilayer Perceptron classifier yielded a sensitivity of 81.3%, a specificity of 81.1%, a recall of 0.81, F-measure 0.81, and an area under the curve (AUC) in receiver operating characteristics (ROC) of 0.836 (Fig. 2) in distinguishing primary brain tumors (glial tumors and glioblastomas) from metastases.

Figure 2
figure 2

Receiver-operating-characteristics (ROC) of the machine learning algorithm to correctly identify primary brain tumors and metastases. The overall area-under-the-curve (AUC) was 0.836.

Full size image

Subanalysis I

In the subanalysis of primary brain tumors, the random forest classifier was able to distinguish glioblastomas from other glial cell tumors (WHO I-III) with a sensitivity of 90.5%, a specificity of 90.4%, a recall of 0.905, F-measure 0.905, and an area under the curve in receiver operating characteristics of 0.868.

Subanalysis II

In the subanalysis of primary brain tumors combined with metastases (Figs. 3, 4), the random forest classifier was able to distinguish these entities with an average sensitivity of 62.5%, a specificity of 74.9%, a recall of 0.625, F-measure 0.628, and an area under the curve in receiver operating characteristics of 0.797 after stratified tenfold cross validation.

Figure 3
figure 3

Cell plot of eight standardized Radiomics features visualize similarities between glioblastomas, other gliomas and metastases. Stark differences can be observed in the values of glioblastomas and metastases.

Full size image
Figure 4
figure 4

Relative visualization of the eight standardized texture analysis features for glioblastomas, other gliomas and metastases. Overall lower values are observed for metastases compared to glioblastomas and other gliomas, respectively.

Full size image

Discussion

In this proof-of-concept study we assessed the utility of radiomics for APTw imaging. To this end, we used radiomics features and machine learning algorithms to differentiate glioblastomas from gliomas and brain metastases. Our results indicate that the application of radiomics to APTw imaging is feasible and allows for the differentiation of these brain neoplasms.

APTw is a novel molecular MRI technique that relies on endogenous cellular proteins in vivo to generate contrast. APTw imaging belongs to the chemical exchange saturation transfer (CEST) imaging group and is the only form of CEST imaging that has yet achieved FDA approval.

The APTw signal is theoretically caused by two major sources: Firstly, the intracellular water-exchangeable amide proton content in the cytoplasm and secondly the base-catalyzed exchange rate at physiological pH range27. Incidentally however, the APTw signal is not pure and may be contaminated by a variety of sources28,29,30,31,32,33. Specifically, the water longitudinal relaxation time (T1) may influence the APTw signal. T1 effects (T1 recovery and T1 related saturation) may influence the APTw signal linearly or in a more complex manner depending on the level of direct water saturation effects, the field strengths of the MR scanner, irradiation power and whether non-steady-state or steady-state acquisitions are performed. Furthermore. The APTw signal may also be affected by semi-solid magnetization-transfer (MT) effects and other nearby CEST and relayed nuclear Overhauser enhancement (rNOE) saturation transfer effects. Ultimately, APTw intensity values may also be impacted by B1 effects which can be triggered by an imperfect distribution of the irradiation power across the brain.

Currently, APTw is mainly used for brain tumor imaging. With malignant brain tumors exhibiting a high degree of protein content, the APTw signal increases steadily with the amount of protein content relative to the surrounding parenchyma. This has been successfully leveraged for differentiating and grading tumors according to their WHO grade34. Specifically, a recent meta-analysis listed the sensitivity and specificity of APTw for differentiating high grade from low-grade glial tumors as 88% and 91% respectively35. Furthermore, based on this principle, high grade regions can be identified within histologically heterogenous brain tumors, thus allowing for more accurate sampling during stereotactic biopsies. Additionally, APTw has also been successfully employed for monitoring tumor response to therapies such as radio-/chemotherapy or high-intensity focused ultrasound. Most importantly, APTw enables the differentiation of recurrent tumor from treatment effects such as radiation necrosis. Lastly, APTw has also been successfully used for identifying genetic markers in gliomas, such as the MGMT or IDH status34.

While APTw has already proven to be a valuable addition to the field of tumor imaging, the development of innovative approaches to further leverage the potential of APTw is highly desirable. Radiomics has become a popular method for extracting more data from radiological images thus enabling in-depth study of the tissue at hand12,36. Therein numerous studies have shown that radiomics approaches based on various imaging modalities (i.e. T1w, T2w, FLAIR, DWI, ADC, SWI, DTI) enhance brain tumor imaging12,16,21.

Here we further enhance the spectrum of radiomics in terms of 3D texture analysis (TA) for brain tumor imaging by applying it to APTw imaging. Our approach yielded a high degree of accuracy in differentiating different types of brain tumors. After dimension reduction all first order TA features were excluded from further analyses due to high intercorrelation. A total of eight features remained for further analysis:

(1) Dependence Non-Uniformity Normalized (DNN), which measures the similarity of voxel dependence throughout the image, an indirect measure of homogeneity. (2) LargeDependenceHighGrayLevelEmphasis (LDHGLE), which measures the joint distribution of large dependence of higher gray-level values. (3) Low Gray Level Emphasis (LGLE), which measures the distribution of low gray-level values, and (4) SumAverage, which measures the relationship between occurences of pairs with lower intensity values and occurences of pairs with higher intensity values. All four aforementioned features are derived from the GrayLevelDependenceMatrix (GLDM), which are used to quantify gray level dependencies in a medical image by taking into account neighbouring voxels. The remaining features were associated with the size of the neoplasms. While VoxelVolume and MeshVolume are directly correlated with size, ZoneVariance measures the variance in zone size volumes. Similarly, ZonePercentage takes into account the number of zones and number of voxels within a defined region of interest, which represents the coarseness of texture. While gliomas were found to exhibit a larger overall size than metastases, which was reflected in these features, they may have no true biological value as metastases and gliomas can present with various sizes.

A major limitation of the current study is its retrospective, single centre design and its small sample size with heterogeneous spectrum of pathologies. It should however be noted, that due to the novelty of the imaging modality (i.e. clinical approval of the sequence occurred in 2018) a larger dataset is not available at the time. Furthermore, larger datasets are likely to decrease the risk of overfitting the machine learning classifiers. We counteracted these limitations by implementing tenfold cross validation of our results. Moreover, the quality of segmentations may strongly influence radiomics results37. Two readers segmented the neoplasms and subsequently features with low interreader agreement were excluded thus reducing bias. However, we acknowledge that there may be more sophisticated methods of performing segmentations and of reducing reader bias. Lastly, our study was performed on a single MR scanner, which limits the generalizability of our findings. The influence of different acquisition parameters, sequences, field strengths, and MR scanners on 3D TA features in APTw images remains to be investigated.

In conclusion, we show that radiomics allows for the differentiation of various brain neoplasms on APTw images. The current work justifies the further study and development of radiomics for APTw imaging in an effort to widen the applicability and utility of APTw imaging for various diseases and anatomies.