The explainability of radiomic-based machine learning models for brain glioma grading on amide proton transfer-weighted images

Abstract

This study aimed to evaluate radiomic-based machine learning models for glioma grading on amide proton transfer weighted (APTw) images using explainability algorithms. A total of 102 patients who underwent preoperative MR examinations, including FLAIR, T1-weighted, T1-weighted contrast-enhanced, and APTw images, were included. Two groups of APTw images were analyzed: one corresponding to contrast-enhanced regions of gliomas and the other corresponding to both contrast-enhanced and peritumoral edematous regions of gliomas. Radiomic features were extracted from these regions. Random forest, support vector machine, naïve bayes classifier, and logistic regression models were trained to distinguish grade 4 from non-grade 4 gliomas. These models were analyzed by Shapley values, permutation importance, and the method of anchors. The results of model explainability analysis revealed that the grading performance of these models relied on radiomic features highlighting the heterogeneity of radiologic phenotypes.

Introduction

The grading of gliomas is critical for the clinical management of patients with brain gliomas. Typically, grading relies on histological samples obtained by surgical resections or biopsies. However, these invasive methods can carry potential risks, such as bleeding, infection. Moreover, due to the intrinsic heterogeneity of brain gliomas, sampling errors are highly probable in the process of sample acquisition. As a result, clinicians strongly desire a non-invasive tool for accurately preoperative determining the grades of brain gliomas.

Recent advancements in the field of radiology have provided opportunities to meet this requirement. The emerging amide proton transfer-weighted (APTw) imaging is a non-invasive MRI technique that has been widely applied in the field of neuroradiology^1,2. Previous radiological studies have demonstrated APTw imaging’s ability to differentiate between high- and low-grade brain gliomas, and highlighted the relatively high grading accuracies across these studies^3–6. Additionally, the correlations between signal intensity of APTw images and glioma grades have also been confirmed by radiological-pathological comparative studies^7,8. Therefore, APTw imaging is a promising non-invasive tool for preoperative grading of brain gliomas.

Among previous radiological studies focusing on glioma grading by APTw images, combining APTw imaging with radiomics methods has emerged as a powerful analytical approach, and it has been validated that integrating radiomics with machine learning algorithms can efficiently improve the grading performance on APTw images^9,10. When compared to other methods, however, a question arises that how machine learning approaches reach such high performance. In fact, because of the black-box nature of many machine learning models, it is often challenging to directly interpret how these models make their conclusions. It is clear that a clinically useful model should be understandable and transparent for human users, and this limitation impacts both the understanding and credibility of model outputs.

Driven by the importance of model transparency, in the field of machine learning, many methods of model explainability have been developed. These approaches can enable a deeper understanding of the intrinsic mechanisms of model behaviors¹¹. As a result, it is feasible that incorporating the model explainability into the pipeline of radiomic study, which facilitating more reliable understanding and application of radiomic models.

In this study, our aim was to gain insights into radiomic-based machine learning models for glioma grading on APTw images by explainability algorithms.

Materials and methods

Patients

In this study, a total of 102 patients were included who underwent preoperative MR examinations and surgical operations in our hospital between March 2021 and September 2022. Inclusion and exclusion criteria were as follows were as follows: (1) Patients underwent preoperative MR scans which included T2-weighted fluid attenuated inversion recovery (FLAIR) images, T1-weighted images, T1-weighted contrast-enhanced images and APTw images; (2) Patients did not receive any treatment before MR scanning; (3) Patients had definitive pathological diagnoses of glioma; (4) Patients were excluded who had concomitant other brain diseases such as cerebral hemorrhage, ischemia or infections; (5) Patients were excluded whose MR image quality was too poor to analyze. Radiological and clinicopathological data were retrieved from the electronic medical records in our hospital. This retrospective study was approved by the ethics committee of our hospital and the informed consent was waived. All methods were performed in accordance with the relevant guidelines and regulations.

MRI acquisition

MRI scans were performed on a 3T Philips Ingenia CX MR scanner. The MR protocol and the scan parameters were briefly listed as follows: the T2-weighted FLAIR sequence (echo time=310ms, repetition time=4800ms, matrix size = 256 × 256, slice thickness= 2 mm), the T1-weighted and contrast-enhanced T1-weighted sequences (echo time = 3.5ms, repetition time = 7.9ms, matrix size = 512 × 512, slice thickness=1 mm), the APTw sequence (echo time = 8.3ms, repetition time=5925ms, matrix size = 256 × 256, slice thickness=6 mm, echo train length = 174, number of averages = 1, saturation power B1 = 2µT, saturation duration=2s).

Delineation of tumoral regions and extraction of radiomic features

The pipeline of radiomic approach in this study was illustrated in Fig. 1.

Fig. 1.

Pipeline of radiomic approach in this study.

Three types of tumor regions were labeled, that is, necrotic tumor core, peritumoral edematous region and enhancing tumor. The definitions of these labels were based on the definitions from Brain Tumor Segmentation (BraTS) challenges 2021¹². The delineation of these labels was performed using the ITK-SNAP software (Version 4.2.0-rc.1) after registration. Every APTw images of every glioma was reviewed and delineated, and the corresponding labels were generated. The delineations were made by two radiologists who were blinded to the grades of gliomas.

Based on these labels, two groups of APTw images were obtained. One was the group containing portions corresponding to the contrast-enhanced regions of gliomas and the other was the one containing portions corresponding to the enhancing tumor and peritumoral edematous regions. The numbers of APTw images in these two groups were 686 and 904, respectively.

Radiomic features were separately extracted from original and filtered APTw images of these groups, including 19 first order statistics features, 16 3D shape-based features, 10 2D shape-based features, 24 Gy level co-occurrence matrix features, 16 Gy level run length matrix features, 16 Gy level size zone matrix features, 5 neighboring gray tone difference matrix features, and 14 Gy level dependence matrix features. The feature extraction was made by the PyRadiomics software package (Version 3.1.0)¹³(https://gitee.com/aaasss0636/xai).

Feature selection, model training and evaluation

The patients were randomly divided into a training set of 65 patients and a test set of 37 patients. As mentioned earlier, due to the two groups of APTw images, this study had two pairs of training and test sets. Random forest, support vector machine, naïve bayes classifier, and logistic regression models were used in this study. The classification objective was to distinguish grade 4 from non-grade 4 gliomas. The model hyperparameter optimization was based on the training sets using the Leave-One-Out-Cross-Validation (LOOCV) method. Feature selection was made before model training and nested inside each LOOCV fold. The Pearson’s correlation coefficient > 0.75 was chosen as the criterion to remove the highly correlated radiomic features. The least absolute shrinkage and selection operator algorithm was used for feature selection. In the LOOCV, the area under receiver operating characteristic curve was used to evaluate the trained models. The final model was retrained the on all of APTw images in every training set. The performance of trained models was assessed on the corresponding test set. All feature selection, model training and evaluation processes were implemented by the Scikit-learn software package (Version 1.4.2)¹⁴.

Model explainability analysis

Shapley values, permutation importance and anchors were separately applied on all types of trained models^15–17. For each of trained models, these methods were used to analyze the model output for each of APTw images from both the training and test sets. The model explainability analysis was implemented by Alibi software package (Version 0.9.5). The kernel SHAP (Shapley additive explanations) algorithm was used to computed the Shapley values. For the radiomic features in every test set, the background or reference data for the kernel SHAP explainer was the corresponding training set. In the method of Shapley values, for each radiomic feature, the mean absolute value of its Shapley value across the training and test sets were separately computed. Permutation importance of every radiomic feature was calculated based on the metric of accuracy. In the method of anchors, a grading rule could be generated from a trained model for a given image. This rule consisted of multiple radiomic features cconnecting by one or more logic AND operators. For each model, a set of grading rules were extracted from all images in a dataset, and the occurrence percentage of every radiomic feature in this set was calculated. Figure 2 illustrated three examples of results from the method of anchors. In the method of anchors, the binning parameters were set as (25, 50, 75).

Fig. 2.

(A-C) Three examples of results from the method of anchors. In every row, from left to right, T1-weighted contrast-enhanced image, T2-weighted fluid attenuated inversion recovery image, amide proton transfer–weighted (APTw) image, and the grading rule generated by the approach of anchors. (A) a 44-year-old man with grade 3 glioma. A support vector machine model, which was trained on the regions of APTw images corresponding to contrast-enhanced regions of gliomas, gave its prediction as non-grade 4. (B) a 62-year-old woman with grade 4 glioma. A support vector machine model, which was trained on the APTw images corresponding to contrast-enhanced and peritumoral edematous regions of gliomas, predicted it was a glioma of grade 4. (C) a 51-year-old man with grade 3 glioma. A random forest model, which was trained on the regions of APTw images corresponding to contrast-enhanced and edematous regions of gliomas, output its prediction as non-grade 4.

Statistical analysis

Quantitative data were presented as mean±standard deviation, and categorical data were expressed as the number of cases. The Mann–Whitney U test was used to analyze differences between two groups of data. Qualitative variables were compared using the Chi-square or Fisher’s exact tests, as appropriate. A p < 0.05 was considered statistically significant. All statistical analyses were performed in R programming language (Version 4.4.1).

Results

Patient characteristics

The age was 48.99 ± 13.80 (range: 4–74) yrs. for all patients in this study. Among them, the ages of the two patients under 18 yrs. were 4 and 12 yrs., respectively. The diameter of gliomas was 72.19 ± 27.39 (range: 18.28–149.80) mm. As showed in Table 1, there were not significantly difference for age, gender, WHO grades and size of lesions between the training and test sets.

Table 1.

demographic characteristics of patients with brain gliomas.

	Training set	Test set	P value
Age (yrs.)	48.02 ± 14.38	50.70 ± 12.75	0.416
Gender: Male/ Female	40/25	21/16	0.250
WHO grades:2/3/4	18/11/36	13/7/17	0.642
Diameter of lesions (mm)	68.98 ± 27.47	77.82 ± 26.68	0.825

Grading performance of models based on radiomic features

Table 2 illustrated the grading capabilities of four machine learning models. Models trained on segmented APTw images corresponding to contrast-enhanced areas achieved slightly higher test accuracies and ROC-AUC.

Table 2.

Grading performance of radiomic machine learning models trained on different regions of amide proton transfer–weighted images of brain gliomas.

	Sensitivity		Specificity		Accuracy		F1-score		ROC_AUC		Brier score
	Enhancing tumor regions	Enhancing tumor and peritumoral edematous regions	Enhancing tumor regions	Enhancing tumor and peritumoral edematous regions	Enhancing tumor regions	Enhancing tumor and peritumoral edematous regions	Enhancing tumor regions	Enhancing tumor and peritumoral edematous regions	Enhancing tumor regions	Enhancing tumor and peritumoral edematous regions	Enhancing tumor regions	Enhancing tumor and peritumoral edematous regions
Random Forest	0.89 95%CI: (0.83, 0.93)	0.79 95%CI: (0.72, 0.84)	0.83 95%CI: (0.76, 0.88)	0.87 95%CI: (0.80, 0.91)	0.86 95%CI: (0.82, 0.89)	0.83 95%CI: (0.78, 0.86)	0.87 95%CI: (0.82, 0.94)	0.85 95%CI: (0.79, 0.90)	0.86 95%CI: (0.85, 0.88)	0.83 95%CI: (0.82, 0.84)	0.15 95%CI: (0.14, 0.17)	0.13 95%CI: (0.12, 0.14)
Support Vector Machine	0.80 95%CI: (0.72, 0.85)	0.72 95%CI: (0.64, 0.78)	0.92 95%CI: (0.87, 0.95)	0.88 95%CI: (0.82, 0.92)	0.86 95%CI: (0.82, 0.89)	0.80 95%CI: (0.75, 0.84)	0.88 95%CI: (0.82, 0.92)	0.83 95%CI: (0.78, 0.86)	0.86 95%CI: (0.84, 0.87)	0.80 95%CI: (0.78, 0.81)	0.16 95%CI: (0.15, 0.17)	0.13 95%CI: (0.12, 0.15)
Naïve Bayes Classifier	0.83 95%CI: (0.77, 0.88)	0.77 95%CI: (0.72, 0.85)	0.67 95%CI: (0.59, 0.73)	0.53 95%CI: (0.45, 0.61)	0.75 95%CI: (0.70, 0.79)	0.65 95%CI: (0.60, 0.71)	0.78 95%CI: (0.70, 0.84)	0.71 95%CI: (0.63, 0.75)	0.74 95%CI: (0.73, 0.75)	0.64 95%CI: (0.63, 0.65)	0.19 95%CI: (0.17, 0.21)	0.20 95%CI: (0.19, 0.21)
Logistic Regression	0.76 95%CI: (0.69, 0.82)	0.40 95%CI: (0.33, 0.48)	0.82 95%CI: (0.75, 0.87)	0.85 95%CI: (0.78, 0.89)	0.79 95%CI: (0.74, 0.83)	0.63 95%CI: (0.57, 0.68)	0.81 95%CI: (0.74, 0.85)	0.66 95%CI: (0.61, 0.69)	0.80 95%CI: (0.79, 0.83)	0.63 95%CI: (0.62, 0.65)	0.16 95%CI: (0.15, 0.17)	0.19 95%CI: (0.18, 0.20)

ROC_AUC: area under receiver operating characteristic curve.

Explainability of machine learning models

For the models analyzed, two types of radiomic features were identified as being associated with glioma grading: the original_shape2D_Sphericity feature and radiomic features related to the inhomogeneity of APTw signal intensity, such as wavelet-HH_glszm_GrayLevelNonUniformity. Furthermore, the importance of these two types of radiomic features could vary considerably depending on the regions of APTw images used for model training (Figs. 3, 4, 5, 6 and 7).

Fig. 3.

Shapley values for machine learning models trained on amide proton transfer–weighted images corresponding to the contrast-enhanced regions of gliomas. In every subplot, the radiomic features are plotted on the vertical axis, sorted by their mean absolute Shapley value, while the horizontal axis shows the corresponding average absolute Shapley value for each radiomic feature.

Fig. 4.

Shapley values for machine learning models trained on the amide proton transfer–weighted images corresponding to contrast-enhanced and peritumoral edematous regions of gliomas. In every subplot, the radiomic features are plotted on the vertical axis, sorted by their mean absolute Shapley value, while the horizontal axis shows the corresponding average absolute Shapley value for each radiomic feature.

Fig. 5.

Results of permutation importance for machine learning models trained on the amide proton transfer–weighted images corresponding to the contrast-enhanced regions of gliomas. In each subplot, the vertical axis lists the radiomic features ranked by the ratio of original to permuted accuracy, and the horizontal axis displays the corresponding ratio value for each radiomic feature.

Fig. 6.

Results of permutation importance for machine learning models trained on the amide proton transfer–weighted images corresponding to contrast-enhanced and peritumoral edematous regions of gliomas. In each subplot, the vertical axis lists the radiomic features ranked by the ratio of original to permuted accuracy, and the horizontal axis displays the corresponding ratio value for each radiomic feature.

Fig. 7.

Occurrence percentages of radiomic features in the method of anchors for machine learning models. (A) the occurrence percentages for models trained on amide proton transfer–weighted images corresponding to the contrast-enhanced regions of gliomas. (B) the occurrence percentages for models trained on the amide proton transfer–weighted images corresponding to contrast-enhanced and peritumoral edematous regions of gliomas.

When models were trained on the APTw images corresponding to contrast-enhanced regions, all explainability methods showed that the original_shape2D_Sphericity played a primary role in grading for both naïve bayes classifier and support vector machine models (Figs. 3, 5 and 7A). In contrast, when these models were trained on the APTw images corresponding to enhancing tumor and peritumoral edematous regions, the original_shape2D_Sphericity held a relatively minor position, while features associated with the inhomogeneity of APTw signal intensity exerted a more significant influence for naïve bayes classifier and support vector machine models (Figs. 4, 6 and 7B).

A similar pattern was also observed in logistic regression and random forest models. The original_shape2D_Sphericity ranked first among all radiomic features for logistic regression and random forest models trained on the APTw images corresponding to enhancing tumor and peritumoral edematous regions (Figs. 4, 6 and 7B). Conversely, radiomic features about signal non-uniformity exhibited relatively high effects for logistic regression and random forest models trained on the APT images corresponding to contrast-enhanced regions (Figs. 3, 5 and 7A).

Discussion

In this study, we evaluated the capacity of various machine learning models to grade gliomas on APTw images and explored insights into how these models make predictions by different explainability techniques. We found that these models relied on radiomic features reflecting the anisotropic nature of tumor shape and the heterogeneity of tumor signal intensity to grade brain gliomas.

Our results indicated that while the underlying principles of these explainability methods differed, an observed commonality was that, models trained on APTw images corresponding to contrast-enhanced regions achieved slightly higher test ROC-AUC/accuracy compared to those trained on enhancing tumor and peritumoral edematous regions. We had a hypothesis about it. We hypothesized that the contrast-enhanced tumoral regions had specific microenvironment compared to other regions. Due to the delineation of ROIs excluded areas of edema and necrosis from these regions, these contrast-enhanced tumoral regions could be thought of a more accurate representation of pure tumor parenchyma, which had fewer effects from non-tumor components. Besides, previous researches had shown that acidic extracellular microenvironments were detrimental to amide proton transfer. Specifically, lower pH levels corresponded to slower amide proton transfer rates, resulting in reduced APTw signal intensities¹⁸. For the contrast-enhanced regions, the increased contrast suggested that these areas were accessible to blood circulation, and resided in a relatively non-hypoxic microenvironment. This microenvironment might facilitate the increase of APTw signal intensities and aid in tumoral area delineation.

The high-throughput nature of radiomic features made them inherently suitable for integrating machine learning techniques into clinical applications. While these models were powerful, their reliability were rarely discussed, as many models lacked transparency for human users. In this study, we incorporated a critical step—model explainability—into our radiomics workflow to gain deeper insights without requiring additional data. To achieve this, we chose three distinct explainability approaches: Shapley values, permutation importance, and anchors. Shapley values were a local method based on game theory that calculated the contribution of each feature to model predictions¹⁵. In this study, for each radiomic feature, the mean absolute value of Shapley value was used, in order to provide a global explanation for the feature importance in model outputs. Similar, permutation importance was a global approach that gauged the importance of input features by assessing how much model performance degraded when feature values were permuted¹⁶. In this study, we calculated the importance of each radiomic feature using the accuracy score and presented it as the ratio between the original score and the permuted score. Anchors, on the other hand, was a local method that explained model behavior by identifying a set of features sufficient to maintain model predictions unchanged¹⁷. This method directly output a set of radiomic features by which models produced predictions. In this study, we considered the occurrence frequencies of radiomic features in the set of rules generated by the method of anchors was correlated with feature importance. Simply put, a higher occurrence frequency suggested greater importance in the process of model decision. In the end, although these explainability methods stemmed from different theoretical frameworks, they all yielded similar conclusions regarding the key radiomic features influencing glioma grading.

Our finding demonstrated that the machine learning models trained on APTw images utilized radiomic features describing tumor heterogeneity to grade brain gliomas. According to the definitions, original_shape2D_Sphericity was a radiomic feature characterizing the shape of lesions, emphasizing the anisotropic and uneven geometric properties of lesion borders. Meanwhile, radiomic features, such as wavelet-HH_glszm_GrayLevelNonUniformity, described the distribution pattern of APTw signal intensity, capturing the non-uniformity of signal distribution in lesions, regardless of whether filtration processing was applied. In this study, all of three explainability methods consistently demonstrated that both types of radiomic features were used in the trained models. Moreover, a common characteristic of these two kinds of radiomic features was their calculations were based on the entire regions demarcated by the selected boundaries, which suggested that they served as global descriptors of lesions. In summary, our results show that, in this study, the glioma grading of machine learning models was based on the radiomic features which highlighted the anisotropic nature of tumor shape and the heterogeneity of APTw signal intensity distribution in tumors.

The limitations of this study should be acknowledged. First, there was an imbalance in the proportions of gliomas, characterized by a lack of cases for grade 1 gliomas and relatively high proportions of grade 4 gliomas. To mitigate it and ensure balanced data distribution for model training, grades 2 and 3 gliomas were combined into a single category. As a result, this study focused on evaluating the ability to distinguish between grade 4 and non-grade 4 gliomas. While this approach simplified the classification problem, it remained clinically significant due to the importance of accurately identifying grade 4 gliomas. Second, as this was a retrospective study, the glioma classification was not based on the recent World Health Organization 2021 classification of central nervous system tumors. However, a recent study demonstrated that the change of glioma classification version did not affect predicting tumor grades by using APTw images¹⁹. Third, in this study, the slice thickness of APTw images was 6 mm, which suggested that that the resolution was relatively low along the Z-axis. This relatively low resolution may compromise the accuracies of 3D shape features. However, we notice that for gliomas, the lesions are usually large, and have complex morphologies. Therefore, we believe that calculating only 2D shape features can result in significant loss of information. Non-isotropic voxels may cause inaccuracies in absolute numerical values for certain features, however, since we are primarily concerned with the distribution patterns of voxels, we believe this limitation does not affect our research objectives. Fourth, this was a single-center study, and the reproducibility and generalizability of explainability methods for radiomic models need to be validated in multi-center studies to ensure broader applicability.

Conclusion

Radiomic-based machine learning models could be used to grade gliomas on APTw images, and the explainability approaches indicated that their grading abilities relied on the radiomic features highlighting the heterogeneity of radiologic phenotypes.

Author contributions

Xuan Gao conceived this study and participated in the literature search, study design, data analysis, data interpretation and wrote the manuscript. Jing Wang participated in data collection and provided critical revision. The authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No.82371945).

Data availability

The data used in this study are available from the corresponding author on reasonable request and with permission of Union Hospital, Tongji Medical College, Huazhong University of Science and Technology.

Declarations

Competing interests

The authors declare no competing interests.