Prediction of Early Recurrence of Adult-type Diffuse Gliomas Following Radiotherapy Using Multi-Modal Magnetic Resonance Images

Abstract

Background:

Adult-type diffuse gliomas are among the central nervous system’s most aggressive malignant primary neoplasms. Despite advancements in systemic therapies and technological improvements in radiation oncology treatment delivery, the survival outcome for these patients remains poor. Fast and accurate assessment of tumor response to oncologic treatments is crucial, as it can enable the early detection of recurrent or refractory gliomas, thereby allowing timely intervention with life-prolonging salvage therapies.

Purpose:

Radiomics is a developing field with great potential to improve medical image interpretation. This study aims to apply a radiomics-based predictive model for classifying response to radiotherapy within the first three months post-treatment.

Methods:

95 patients were selected from the Burdenko Glioblastoma Progression Dataset. Tumor regions were delineated in the axial plane on contrast-enhanced T1(CE T1W) and T2 fluid-attenuated inversion recovery (T2_FLAIR) magnetic resonance imaging. Hand-crafted radiomic (HCR) features including first- and second-order features were extracted using PyRadiomics (3.7.6) in Python (3.10). Then, recursive feature elimination with a random forest (RF) classifier was applied for feature dimensionality reduction. RF and support vector machine (SVM) classifiers were built to predict treatment outcomes using the selected features. Leave-one-out cross-validation was employed for tuning hyperparameters and evaluating the models.

Results:

For each segmented target, a total of 186 hand-crafted radiomics features were extracted from the MRI sequence. Using the top-ranked radiomic features from a combination of CE T1W and T2_FLAIR, an optimized classifier achieved the highest averaged AUC (area under the curve) of 0.829±0.075 using the RF classifier. The HCR features of CE T1W produced the worst outcomes among all models (0.603±0.024 and 0.615±0.075 for RF and SVM classifiers respectively).

Conclusions:

We developed and evaluated a radiomics-based predictive model for early tumor response to radiotherapy, demonstrating excellent performance supported by high AUC values. This model, harnessing radiomic features from multi-modal MRI, showed superior predictive performance compared to single-modal MRI approaches. These results underscore the potential of radiomics in clinical decision support for this disease process.

1. Introduction

The most common aggressive malignant primary tumors of the central nervous system (CNS) are adult-type diffuse gliomas^1,2. The standard treatment regimen for gliomas includes maximal safe surgical resection followed by adjuvant radiotherapy (RT) and concurrent temozolomide chemotherapy³. Despite significant advancements in neurosurgical, chemo, and radiotherapeutic treatments, prognosis remains dismal^3–5, with a local recurrence rate of 75% to 80%, which indicates a poor survival outlook^6,7. The two-year survival rate for these patients is low, and less than 5% survive beyond five years post-diagnosis⁴. Therefore, accurate and timely detection of recurrent or refractory disease is critical to improving survival outcomes by promoting timely therapeutic interventions⁸. Historically, tumor response has been monitored through serial Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) scans, assessed by physicians relying heavily on their clinical acumen⁹. However, this manual evaluation is inherently subjective and time-consuming, with a potential risk of misdiagnosis. While biopsy remains the gold standard for confirming tumor recurrence, it is invasive, expensive, and not without medical risk¹⁰. An alternative method can be machine learning (ML) with radiomics which has gained increasing attention in medical imaging and neuro-oncology. Previous studies have demonstrated that radiomics can be used to quantitatively extract and evaluate numerous imaging features to classify gliomas effectively^11,12 and differentiate tumor recurrence from radiation necrosis^13,14. Therefore, this study focuses on analyzing and extracting quantitative imaging features from regularly obtained imaging data to create predictive or prognostic models based on this information. For this purpose, hand-crafted radiomics (HCR) features were extracted from two MRI sequences to assess the treatment response in gliomas. Models were built for each MRI sequence and a combination of both to evaluate the effectiveness of using multimodal MRIs. Furthermore, we propose a robust algorithm based on an optimally selected feature subset (constructed with the minimum number of features using recursive feature elimination)¹⁵ and a Synthetic Minority Over-sampling Technique (SMOTE)¹⁶ to address the class imbalance and improve classification performance. By developing the prediction model, we believe that early recurrence can be detected in patients at risk within three months after completing their treatment.

2. Related works

Automated image processing methods, such as radiomics with ML, have demonstrated effectiveness in differentiating tumors, and therefore have the potential to replace traditional approaches. ^10,17–19. For example, Patel et al²⁰ utilized radiomic features extracted from MR images with clinical (age) and molecular (MGMT promoter methylation status) data to differentiate between early true progression and pseudoprogression in GBM patients. They used two shape-based features to enhance the disease mask (elongation and sphericity), three apparent diffusion coefficient or ADC radiomic features from the enhancing disease mask (kurtosis, correlation [GLCM], and contrast [NGTDM]), and one T2-weighted (T2W) radiomic feature from the perilesional edema mask (dependence entropy [GLDM]). Their final model achieved an area under the receiver operating characteristic curve (AUC) of 0.80. In addition, the model obtained a sensitivity of 78.2%, a specificity of 66.7%, and an accuracy of 73.7%. Lohmann et al²¹ explored the potential of using O-(2-[18F]fluoroethyl)-L-tyrosine (FET) positron emission tomography (PET) radiomics to correctly identify patients with pseudoprogression. The study calculated 944 features on unfiltered and filtered images of 34 patients who showed progressive MRI changes within the first 12 weeks after completing temozolomide chemoradiation. The highest achieved diagnostic accuracy was 79% with a sensitivity of 69% and specificity of 89%. A recent study to predict the early recurrence of glioblastoma following surgery used a combination of accessible Rembrandt image features of glioblastoma, radiomics features from preoperative brain MR images and clinical data. Patients were excluded who had missing clinical data or received treatment before surgery, such as neoadjuvant radiotherapy, chemotherapy, or chemoradiotherapy. The model successfully distinguished between patients with early recurrence and those with later recurrence. In the training cohort, the model achieved an AUC of 0.85 (95% CI, 0.77–0.94), and in the validation cohort, it achieved an AUC of 0.84 (95% CI, 0.71–0.97) ²². However, the model was built based on 2D features extracted from the brain MRIs. Shim et al²³ developed a radiomics-based neural network to predict recurrence patterns (local recurrence vs distant recurrence) in patients with GBM. They extracted perfusion radiomic features from contrast-enhanced weighted imaging (CE T1W) and T2_FLAIR sequences based on the cerebral blood volume (CBV) maps as potential biomarkers to predict the prognosis of GBM. The authors recorded recurrences at one year to create binary outcomes in order to simplify the prediction model. Additionally, each model was built using each MRI sequence separately. Further, the study lacked comparisons with the original data to show the impact of wavelet filters on the model’s performance. The key findings of the study are AUC of 0.969 and 0.864 for local and distant recurrence, respectively. Liu et al¹⁹ developed a radiomics prediction model to differentiate between recurrence and non-recurrence using pre-radiotherapy and recurrence brain MRIs (T1W, CE T1W, and T2W) of patients with gliomas. The AUC of the patient model on the validation dataset was 0.648 (0.474–0.821), 0.593 (0.398–0.787), and 0.693 (0.515–0.871) for T1W, T2W, and CE T1W respectively. However, this study did not use any filter, and similar to the study by Shim et al., did not combine radiomics features from different MRI sequences.

Tumors are composed of a variety of biological tissues, and a single MRI sequence may not provide a comprehensive view of abnormal tissues. By combining different complementary sources of information, we can improve the accuracy of predictions and ultimately enhance outcomes for patients. Additionally, GBM is a highly vascularized tumor, and the degree of vascularization can directly impact the prognosis and recurrence of GBM patients ²⁴. Therefore, it is reasonable to assume using different MRI sequences, which take into account various imaging features including enhancing characteristics that reflect tumor vascularity, could serve as a valuable tool for predicting the early recurrence of GBM ²⁵. The findings from Wang et al²⁶ study revealed superior model performance when combining features from all MRI sequences with an AUC of 0.966 and 0.930 for training and validation data sets respectively. Their model was able to make a prediction for only grade II glioma in less than one year after the first resection.

As a result, this study aims to achieve the following goals: First, early images (≤3 months) assessment to predict the tumor response to chemoradiation treatment using multi-modal MR. Second, to determine the significance of using multimodal vs single modal MRIs using hand-crafted radiomics (HCR) features. Third, assess the effectiveness of edge enhancement filters on the prediction performance. We believe the results of this study have future implications for treatment strategies for glioma in clinical practice.

3. Material & Methods

Our method was constructed using contrast-enhanced T1 (CE T1W) and T2 Fluid Attenuated Inversion Recovery (T2_FLAIR), the preferred sequences for characterizing glioma due to their high contrast and spatial resolution. While these images provide valuable information about the size and location of a brain tumor, they are frequently insufficient to definitively distinguish between tumor recurrence and radiation necrosis. This motivates us to develop and evaluate a radiomics method coupled with ML to enhance the sensitivity and specificity of MR imaging. Radiomics is based on the hypothesis that disease-induced histopathological and genetic changes manifest as distinct phenotypes detectable in medical imaging. Utilizing this approach allows for objectively and quantitatively extracting rich features from the images, significantly contributing to patient care²⁷.

3.1. Preprocessing

This study utilized the Burdenko Glioblastoma Progression Dataset (BGPD), a comprehensive collection of images from 180 patients diagnosed with gliomas and treated at the Burdenko National Medical Research Center of Neurosurgery from 2014 to 2020^28,29. The dataset contains information on the status of IDH1/2 gene mutation and MGMT promoter methylation. According to the latest edition of the WHO (World Health Organization) classification of CNS tumors (CNS WHO 2021), adult-type diffuse gliomas are classified into three tumor types: Astrocytoma, IDH-mutant; Oligodendroglioma, IDH-mutant and 1p/19q-codeleted; Glioblastoma, IDH-wildtype 1. Thus, we considered this database as adult-type diffuse gliomas, and it comprises imaging studies conducted for radiotherapy planning with one to eight follow-ups per patient. The radiotherapy studies involve 4 MRI sequences (T1W, CE T1W, T2W, FLAIR), CT scans, and related radiotherapy planning files (RTSTRUCT, RTPLAN, and RTDOSE). The MRI studies were obtained from different sites using scanners (Siemens, Philips, and General Electric) with different scanning protocols. The initial pixel size was 256 × 256 or 512 × 512 and thickness was variable from 0.8 cm to 5.0 mm ³⁰. For each patient, the follow-up studies include at least a set of CE T1W and T2_FLAIR examinations. After carefully examining images to ensure an accurate assessment of tumor response to radiation therapy, patients with less than three follow-ups and MR images with layer gap > 5 mm and obvious artifacts were excluded. The study comprised 95 patients (49 male, 46 female) with an average age of 57 years and an age range of 18 to 79 years. For the early evaluation of glioma post-treatment, we utilized MRI scans acquired one to two months after completing radiation therapy. The contours available in these images were used for feature extractions. In cases where the images lacked RT structures, pre- and post-images were fused using Velocity (Ver. 4.1). A thorough slice-by-slice examination was conducted by a skilled Radiation Oncologist. Following the examination, the expert physician performed the delineation of tumors independently, blinded to the histological results in 3D Slicer software (Ver 5.6.1)^31,32 on the axial plane of CE T1W and T2_FLAIR. Normalization with the final 256 bins was performed on all original MR images using the gray-scale discretization method before extracting the HCR features to remove the potential differences between MRIs acquired from the different MRI scanners. Moreover, to homogenize image processing, we performed resampling on both the CE T1W (0.5 × 0.5 × 0.5 mm) and T2_FLAIR (0.5 × 0.5 × 0.5 mm). The processed brain MR images were subjected to a three-level wavelet (Haar) ³³ decomposition (Discrete Wavelet Transformation (DWT)) and a gradient filter to eliminate noise. DWT can analyze a signal at various levels of detail. This is achieved by processing the signal through a series of band-pass filters. To accomplish this, the signal is successively filtered through low-pass (L) and high-pass (H) filters³⁴. The 3D wavelet filter generates sub-volumes on three spatial axes (x, y, z), producing an 8 sub-band filter. These sub-bands consist of high-pass filters (HLL, HLH, HHL, HHH) and low-pass filters (LLL, LLH, LHL, LHH). Subsequently, the HCR features were extracted from each of the MRI images.

3.2. Feature extraction

We utilized 3D imaging data to extract features, allowing for more comprehensive and precise feature identification through various operations. 93 HCR features were extracted from each MR imaging sequence using PyRadiomics version 3.7.6³⁵. These features include First-order (n=18), Grey Level Co-occurrence Matrix (GLCM, n=24), Grey Level Size Zone Matrix (GLSZM, n=16), Grey Level Run Length Matrix (GLRLM, n=16), Grey Level Dependence Matrix (GLDM, n=14), and Neighboring Gray Tone Difference Matrix (NGTDM, n=5).

3.3. Feature Selection

After combining each subject’s CE T1W and T2_FLAIR features in a feature vector (FV)³⁶, we employed Recursive Feature Elimination with Cross-Validation (RFE-CV) to develop a trained machine-learning model that minimizes the required features to yield predictions with useful accuracy. The scikit-learn library provides an RFE-CV model for the dimensional reduction of the training dataset³⁷. This approach involves selecting a subset of features by iteratively eliminating less important ones. RFE requires a set number of features to be retained, but often this number is unknown beforehand. Therefore, to determine the optimal number of features, Leave-One-Out cross-validation (LOOCV) was used with RFE to score various feature subsets and select the best-scoring collection of features in the present study. Consequently, this process yielded a ranked list of features and an empirically determined optimal set, enhancing prediction accuracy for the targeted classification problem³⁸. Moreover, removing unnecessary noise from less important features regularizes the model and prevents overfitting³⁹. This study employed RFE with a Random Forest (RF) classifier. It is worth noting that the algorithm used in RFE does not have to be the algorithm that fits the selected features; different algorithms can be used ³⁸. This process was also performed on each MR sequence separately to compare multimodal and single-modal MR classification models. Each group’s RFE-CV was run over 30 iterations to identify and select the most frequently used features.

3.4. Classification

The ML modeling process was implemented in Python (Version 3.10). Feature values were normalized to [0, 1] to improve training stability. The study’s dataset includes the treatment response status for each subset of patients. According to the provided information, this study contained three response classes: stable (n=39 patients), response (n=26), and progression (n=30). An imbalanced dataset results in poor classification performance⁴⁰. To address this issue, we employed SMOTE to generate synthetic samples for the minority class¹⁶. LOOCV was utilized to tune hyperparameters and evaluate performance. This configuration substantially mitigates the risk of overfitting, ensuring the model is robust and optimally tuned. We employed two machine learning models to predict and compare early tumor responses using different methodologies: a Support Vector Classifier (SVC) and a Random Forest Classifier (RFC). Since the aim of detection frameworks is to predict unexpected data, successfully assessing the model’s performance is crucial. The Area Under the Receiver Operating Characteristic Curve (AUC) is reported to illustrate the trade-off between precision and recall and is used to evaluate the performance of the models at different threshold settings⁴¹.

3.5. Statistical Analysis

Correlation is frequently used in the statistical data analysis of radiomic features of medical images. This helps in feature clustering to avoid redundancy. According to the normality of samples based on the Shapiro-Wilk test, our data does not follow a normal distribution. Hence, we employed the Mann-Whitney U-test to compare groups and the Spearman correlation to measure the statistical dependence between the selected radiomic features. The Spearman correlation coefficient was calculated based on Quinnipiac University as follows: The ranges were defined to be 0.0 < CC < 0.2 for no correlation, 0.21 < CC < 0.3 for weak correlation, 0.31 < CC < 0.4 for moderate correlation, 0.41 < CC < 0.7 for strong correlation, and CC greater or equal to 0.7 for very strong correlation ⁴². Correlation matrices were obtained with Python (Ver.3.11.4).

4. Results

We conducted this study using two different brain MRI sequences (CE T1W and T2_FLAIR) for predicting early glioma response to RT (stable, response, and progression). Figure 1 presents representative MR images from the dataset, the impact of the gradient filter, and examples of wavelet filters on CE T1W and T2_FLAIR.

Figure 1.

A close-up of a brain MRI scan from the employed dataset. a1) CE T1, b1) T2_FLAIR. Cropped image to better illustration of the tumor a2) CE T1W, b2) T2_FLAIR. Gradient filter effect on a3) CE T1, b3) T2_FLAIR, wavelet filter on a4) CE T1, b4) T2_FLAIR. The blue line illustrates the region of interest (ROI) delineated by the radiation oncologist for feature extraction. CE T1W refers to contrast-enhanced T1 weighted, and T2_FLAIR stands for T2 Fluid Attenuated Inversion Recovery.

4.1. Feature Selection

This study implemented the RFE method using the LOOCV with a random forest classifier. An RF is an ensemble algorithm of bagging methods that achieves increased prediction performance by creating multiple decision trees for the input data and combining their results²⁰.

Additionally, to validate the effectiveness of using multiparametric features vs single parameter features to classify the response of glioma to the treatment, the feature selection algorithm was separately applied to CE T1W, T2_FLAIR, and a combination of CE T1W and T2_FLAIR. The results of the feature distribution of 8 sub-band wavelets, Gradient, and Original (no-filter) are presented in Table S-1. The shared common features in all three groups (CE T1W, T2_FLAIR, and a combination of CE T1W&T2_FLAIR) were tabulated in Table S-2.

4.2. Model Evaluation

Tables 1 summarize the results of AUC values for CE T1W, T2_FLAIR, and the combination of both MR sequences in detail.

Table 1.

Average AUC or area under the curve values using features extracted from CE T1W, T2_FLAIR, and CE T1W&T2_FLAIR MR images for Random Forest Classification (RFC) and Support Vector Classification (SVC). For each group of MR sequences, the highest average AUC achieved is highlighted in bold. CE T1W refers to contrast-enhanced T1 weighted, and T2_FLAIR stands for T2 Fluid Attenuated Inversion Recovery.

	CE T1W		T2_FLAIR		CE T1W & T2_FLAIR
Filter	RFC	SVC	RFC	SVC	RFC	SVC
Original	0.692±0.079	0.735±0.092	0.656±0.097	0.659±0.071	0.790±0.051	0.740±0.100
Gradient	0.722±0.086	0.641±0.144	0.729±0.032	0.715±0.097	0.765±0.044	0.729±0.075
HHH	0.630±0.10	0.627±0.077	0.780±0.078	0.744±0.032	0.703±0.087	0.686±0.069
HHL	0.603±0.024	0.615±0.075	0.741±0.099	0.772±0.086	0.728±0.082	0.688±0.107
HLH	0.731±0.076	0.673±0.082	0.782±0.044	0.75±0.079	0.809±0.044	0.651±0.105
HLL	0.657±0.054	0.691±0.075	0.758±0.035	0.712±0.074	0.824±0.057	0.790±0.073
LHH	0.730±0.071	0.680±0.052	0.785±0.075	0.749±0.086	0.751±0.064	0.666±0.124
LHL	0.727±0.074	0.658±0.052	0.797±0.066	0.805±0.051	0.829±0.075	0.792±0.101
LLH	0.799±0.037	0.730±0.077	0.739±0.074	0.712±0.081	0.747±0.038	0.757±0.078
LLL	0.814±0.071	0.796±0.089	0.782±0.069	0.759±0.089	0.817±0.047	0.650±0.077
All filters	0.756±0.123	0.773±0.074	0.788±0.08	0.768±0.066	0.823±0.067	0.855±0.057

4.3. Statistical Analysis

Due to the non-normal distribution of our data, we made use of the Spearman correlation to assess the statistical relationship between the specific radiomic features that were selected for analysis. Heatmaps are an effective tool for visualizing data and revealing patterns, relationships, and similarities within datasets. Heatmaps provide even deeper insights into the data in conjunction with hierarchical clustering.

Figure S-1 shows the Spearman correlation heatmaps with Hierarchical Clustering of all filters and the LLL and LHL sub-bands of the wavelet filter for CE T1W, T2_FLAIR, and the combination of both respectively. The Correlation Heatmaps with Hierarchical Clustering of the gradient filter, other sub-bands of wavelet filters, and original groups selected from CE T1W & T2_FLAIR are reported in the Appendix.

Since the data of 95 patients were not normally distributed, the statistical significance of the data between SVC and RFC was tested with the Mann-Whitney U test. It shows that RFC outperforms SVC in all groups, especially using a combination of two MRI sequences (p-value = 0.003). On the other hand, no significant difference was observed between SVC and RFC using CE T1W or T2_FLAIR features (p-value > 0.05). Figure 2 demonstrates the results of this test for LLL of CE T1W, LHL of T2_FLAIR, and a combination of both MR sequences.

Figure 2.

Results of median comparison of area under the curve (AUC) between support vector classification (SVC) and random forest classification (RFC) for a) LLL of Contrast-enhanced T1 (CE T1W), b) LHL of T2_FLAIR, and c) LHL of the combination of CE T1W & T2_FLAIR.

Figure 3 demonstrates the significance of using multimodal MRI compared to single modal images for RFC. This figure shows that the combination of CE T1W and T2_FLAIR increases the AUC value by 7% and 3% compared to CE T1W and T2_FLAIR, respectively. A statistical significance difference was observed between CE T1W and T2-FLAIR, and CE T1W, and the combination of two MRIs (p-value < 0.05). However, there is no significant difference between T2-FLAIR and CE T1W & T2-FLAIR (p-value > 0.05).

Figure 3.

The bar chart shows the median comparison of area under the curve (AUC) between Contrast-enhanced T1 (CE T1W), T2_FLAIR, and a combination of both MR sequences for random forest classification (RFC).

Figure 4 displays three different samples from the dataset: stable, response, and progression cases. It illustrates the tumors’ conditions before starting treatment, a month after ending treatment, and 5 months after completing treatment. For the stable case illustrated in Figure 4, both models produced inaccurate predictions when using LLL of CE T1W HCR features. The RFC model accurately predicted this case using HCR features of T2_FLAIR and a combination of both MR sequences, while the SVC correctly predicted using T2_FLAIR, but inaccurately predicted using features from both MR sequences. Response case: All models produced accurate results utilizing LHL features extracted from CE T1W, T2-FLAIR, as well as a combination of both MR sequences to build. Progression case: The RFC model predicted inaccurately using the LHL features CE T1W and a combination of CE T1W&T2_FLAIR while yielding correct prediction using T2_FLAIR. The SVC correctly classified this case using extracted features from T2_FLAIR, and a combination of both MR sequences. However, it produced an inaccurate prediction when using LHL of CE T1W.

Figure 4.

Illustrates contrast-enhanced T1 weighted (CE T1W) before, 1^st follow-up (a month) after, and third follow-up (5 months) after the treatment for stable, response, and progression cases respectively.

5. Discussion

Early diagnosis is crucial for effective treatment of brain tumors. Artificial intelligence has led to significant advances in this area, particularly through radiomics. This emerging field involves converting medical images into high-dimensional, analyzable features using machine-learning models, thus improving tumor diagnosis, and enhancing medical decision-making. We present a novel method that applies ML with radiomic features for the early prediction of tumor response after radiotherapy. This study has several unique characteristics and generates several new observations.

Our research explored the effectiveness of two different classifiers in conjunction with various imaging modalities, and filters (DWT, and Gradient) leading to useful results. The framework we developed achieved near-perfect AUC values, indicating its high predictive accuracy. Specifically, using Support Vector Classification, the highest averaged AUC values recorded were 0.796±0.089 for CE T1W, 0.805±0.051 for T2_FLAIR, and 0.792±0.101 for their combined use. The Random Forest Classifier showed even more promising results, surpassing the SVC, particularly in combined CE T1W and T2_FLAIR modalities, with the highest mean AUC values of 0.829±0.075.

Moreover, our analysis demonstrated the superior performance of classifiers when applied to multi-modal MRI data compared to single-modal data, as depicted in Figure 3. The application of features extracted from multi-modal MR images enhanced AUC values by 7% and 3% compared to CE T1W or T2_FLAIR alone. CE T1W imaging is commonly used to determine the gross tumor volume, but it may not show the entire extent of the tumor. In addition to CE T1W, FLAIR imaging is also used to delineate glioblastoma tumor volumes in MRI. Hyperintense T2-FLAIR volumes are used to pinpoint areas of cancer infiltration in edema, particularly in radiotherapy treatment planning⁴³. However, the prognostic value of the FLAIR signal has not been clearly established in assessing glioblastoma treatment response⁴⁴. Combining these sequences enables us to detect different aspects of tumor biology with high sensitivity. These include blood-brain barrier breakdown, necrosis, edema, non-enhancing tumor infiltration, and blood flow. In other words, this multimodal approach allows us to better differentiate between different parts of the tumor that likely reflect local cellular phenotypes and genotypes⁴⁵.

To enhance the prediction of early response to radiotherapy, we strategically employed recursive feature elimination with cross-validation, focusing on the selection of robust, high-performing features. One of the most important steps in radiomic studies lies in effective feature selection⁴⁶, which directly influences both the accuracy and robustness of the ML models. Various feature selection methods were used in previous glioma studies for dimensionality reduction based on the wrapper and filtration approaches^11,47–49. However, some radiomic features were found useful in classification, while others were shown not to be significantly related^50–52. Therefore, we try to identify the most effective features by utilizing different parametric combinations of various MR sequences, instead of relying on a single sequence^53,54. Our study extracted HCR from CE T1W and T2_FLAIR, and then RFE with cross-validation was applied to select the most relevant subset by removing unnecessary features. The main logic behind RFE is that the most relevant features will have the highest impact on the target variable, and thus will be more useful for predicting the target category⁴⁹. It can be used in combination with any model to score different subsets of features to finally select the best one³⁸. This study implemented the RFE method using the LOOCV combined with the RF classifier for feature reduction^15,20. We ran this model over 30 iterations to obtain the frequency of occurrence of each feature. We then retained a subset of radiomic features with the highest frequency across the cross-validation runs for further analysis (Table S-2). The selected features were then used to develop a robust model to classify the early response of gliomas to chemoradiation by applying two different ML classifiers in the test cohort.

Unlike prior studies, we examined the impact of different image preprocessing filters on the model’s classification performance. The average AUC values for RFC and SVC classifiers using features extracted from CE T1W, T2_FLAIR, and a combination of both MRI sequences can be found in Table 1. The results exhibit distinct performance differences based on the MRI modification and the type of filter applied. Table 1 demonstrates that applying edge enhancement filters can outperform the outcome of the classifications. Among the 3D wavelet filters, low-pass filter groups (LLH, LHL, LHH, LLL) provided better performance among all filter combinations. The results correspond to the original expectation that there will be noticeable distinctions between the original, gradient, and wavelet filters. This is because the wavelet filter is capable of analyzing multifrequency phenomena localized in space. By splitting a signal into different frequency sub-bands, various types of signals containing different frequency characteristics can be extracted from MR images. Hence, the essential features that signify whether a tumor is benign or malignant are likely identifiable within a certain frequency sub-band¹³. For CE T1W, the highest AUCs were achieved with the LLL filter for both RFC (0.814±0.071) and SVC (0.796±0.089), suggesting that low-frequency components effectively capture significant predictive information from CE T1W images. In contrast, T2_FLAIR showed the best results with the LHL filter for both RFC (0.797±0.066) and SVC (0.805±0.051), indicating its robustness in providing features essential for predicting glioma recurrence.

Furthermore, the combination of CE T1W and T2_FLAIR MRI types yielded the best overall performance when processed with the LHL filter, achieving the highest AUC for RFC (0.829±0.075) and very strong results for SVC (0.792±0.101). This superior performance underscores the benefits of integrating features from both MRI modalities, which enhances the classifier’s ability to predict early recurrence effectively. Additionally, when all filters were combined, the AUC values were consistently high across both MRI types and their combination, with the combined modalities showing the highest AUC for SVC (0.855±0.057). This suggests that a comprehensive feature set from all filters provides a robust representation for classifying glioma responses, demonstrating the potential of multi-modal MRIs and diverse radiomic features to enhance prediction accuracy for early tumor response to treatments.

We further examined the impact of different feature groups on the model’s classification performance. The top 10 features extracted from CE T1W, T2_FLAIR, and their combination exhibit distinct characteristics. Figures S-1(a), S-1(b), and S-1(c) focus on specific wavelet sub-bands. In contrast, Figures S-1(d), S-1(e), and S-1(f) include all filters, capturing a broader set of features for each modality and their combination. CE T1W features, such as Skewness, Uniformity, and Median, derived from first-order statistics, GLCM, and NGTDM, show high correlations in Figure S1(a) due to enhanced tissue contrast. In comparison, the T2_FLAIR features, including Difference Entropy, Mean, and Size Zone Non-Uniformity Normalized, display varied correlations in Figure S-1(b), capturing lesion heterogeneity. The combination of CE T1W and T2_FLAIR in Figure S1(c) yields a richer feature set, leveraging the strengths of both modalities to provide a more comprehensive representation. Spearman correlation heatmaps in Figures S-1(d), S-1(e), and S-1(f) further illustrate these differences. CE T1W features in Figure S-1(d) show tight clustering and high correlations, reflecting strong internal consistency in capturing tissue contrast variations. T2_FLAIR features in Figure S-1(e) exhibit varied correlation patterns, capturing a broader range of lesion-related information. The combined features in Figure S-1(f) present a complex correlation structure, demonstrating the integration of comprehensive information from both modalities, which is crucial for accurate glioma recurrence prediction.

As illustrated in Table S-2, first-order, NGTDM, and GLDM features have the least impact on classifications among all features. On the other hand, GLCM, GLRLM, and GLSZM are the most frequently selected features for building high-performing ML models. This finding aligns with existing studies that highlight the effectiveness of GLCM, GLRLM, and GLSZM in detecting minor pixel intensity changes, adeptly distinguishing between cancerous and healthy tissues by focusing on the spatial positioning of pixels^51,55,56. For instance, studies by Amin et al.¹⁷ and Kibriya et al.³⁶ have successfully employed GLCM in brain tumor detection, achieving high accuracies of up to 99% with SVM and KNN classifiers respectively. Molina et al⁵⁷ investigated the robustness of GLCM and GLRLM under dynamic range and spatial resolution changes in 3D MRI of the brain. They pointed out that GLRLM provides a more accurate understanding of tumor heterogeneity as it can identify complex 3D structures that share the same grey-level values. Additionally, Mohanty et al.’s⁵⁸ research using these features in mammogram analysis further confirms their effectiveness in distinguishing between benign and malignant masses with an accuracy of 92.3%. GLSZM is an extension of GLRLM into a higher dimension and its matrix is calculated based on the RLM (run length matrix). Accordingly, it is expected that the two matrices contain similar heterogeneity information since they are highly likely to be correlated⁵⁶.

Moreover, our analysis of the data in Table S-2 identified skewness (HHL, HLL, LHH, LHL, LLH) among the first-order features as the most influential feature in the early prediction of glioma treatment response. For GLCM, six features (correlation, cluster shade, Idn, Imc1, Imc2, and MCC), 2 features in GLRLM (gray level non-uniformity and Run length non-uniformity), 3 features in GLSZM (large area high gray level emphasis, size zone non-uniformity, and large area high gray level emphasis), 3 features in GLDM (dependence non-uniformity, dependence variance, and gray level non-uniformity), and coarseness feature in NGTDM are found to be more reproducible, robust for all conditions and create significant differences between the three classification groups. This study underscores the utility of radiomics and machine learning to predict the early response of glioma to radiotherapy and aid in clinical decision-making. Despite encouraging results, we acknowledge that the study has limitations. Our dataset is small and contains only two MR sequences (CE T1W and T2_FLAIR), which can limit the generalizability of our study results. Also, we could not evaluate the overall survival due to incomplete clinical data and the lack of long-term follow-up for some patients. A statistical correlation was not applied to select features after feature reduction, which can introduce bias into the analysis. To validate the clinical use of radiomic models for assessing treatment response in glioma patients and to improve the robustness and generalizability of the model, further research is needed with larger multi-institutional datasets, including molecular markers, additional imaging studies, and more comprehensive clinical outcomes data. We believe these models can serve as valuable tools for making personalized treatment decisions and ultimately enhancing the quality of life for glioma patients.

6. Conclusion

This paper presents a highly accurate scheme integrating radiomics and machine learning with multimodal MR image data to predict early glioma response to radiotherapy. Our approach utilizes radiomic features from multimodal MRI scans, outperforming traditional single-modal MRI techniques in predictive capability. We enhanced model accuracy by applying edge enhancement filters in preprocessing to isolate the tumor from surrounding tissue. The encouraging results suggest a significant role for radiomics in personalizing patient treatment plans and assessing new therapy efficacy. Future research should aim to corroborate these findings with new, large, and diverse imaging datasets.

Supplementary Material

Supinfo1

Figure S-1. Spearman correlation heatmap with hierarchical clustering of the top 10 selected features with a threshold of 0.8, a) LLL from CE T1W, b) LHL from T2_FLAIR, and c) LHL from the combination of CE T1W and T2_FLAIR. All filters included wavelet & gradient, D) CE T1W, E) T2_FLAIR, F) CE T1W & T2_FLAIR. CE T1W stands for contrast-enhanced T1 weighted, and FLAIR refers to T2 Fluid Attenuated Inversion Recovery. The magnitude of the correlation is illustrated in the color bar on the bottom left.

Figure S-2. Spearman correlation matrices (Heatmap) with hierarchical clustering of top 10 selected features, a) Original, b) Gradient, wavelet filters: (c) HHH, (d) HHL, (e) HLH, (f) HLL, (g) LHH, (h) LLH, (i) LLL from CE T1W & T2_FLAIR. CE T1W and FLAIR refer to contrast-enhanced T1 and T2_FLAIR respectively.