An MR Radiomics Framework for Predicting the Outcome of Stereotactic Radiation Therapy in Brain Metastasis

A Multidimensional Connectomics- and Radiomics-Based Advanced Machine-Learning Framework to Distinguish Radiation Necrosis from True Progression in Brain Metastases

We introduce tumor connectomics, a novel MRI-based complex graph theory framework that describes the intricate network of relationships within the tumor and surrounding tissue, and combine this with multiparametric radiomics (mpRad) in a machine-learning approach to distinguish radiation necrosis (RN) from true progression (TP). Pathologically confirmed cases of RN vs. TP in brain metastases treated with SRS were included from a single institution. The region of interest was manually segmented as the single largest diameter of the T1 post-contrast (T1C) lesion plus the corresponding area of T2 FLAIR hyperintensity. There were 40 mpRad features and 6 connectomics features extracted, as well as 5 clinical and treatment factors. We developed an Integrated Radiomics Informatics System (IRIS) based on an Isomap support vector machine (IsoSVM) model to distinguish TP from RN using leave-one-out cross-validation. Class imbalance was resolved with differential misclassification weighting during model training using the IRIS. In total, 135 lesions in 110 patients were analyzed, including 43 cases (31.9%) of pathologically proven RN and 92 cases (68.1%) of TP. The top-performing connectomics features were three centrality measures of degree, betweenness, and eigenvector centralities. Combining these with the 10 top-performing mpRad features, an optimized IsoSVM model was able to produce a sensitivity of 0.87, specificity of 0.84, AUC-ROC of 0.89 (95% CI: 0.82-0.94), and AUC-PR of 0.94 (95% CI: 0.87-0.97).

A Self-Attention-Guided 3D Deep Residual Network With Big Transfer to Predict Local Failure in Brain Metastasis After Radiotherapy Using Multi-Channel MRI

A noticeable proportion of larger brain metastases (BMs) are not locally controlled after stereotactic radiotherapy, and it may take months before local progression is apparent on standard follow-up imaging. This work proposes and investigates new explainable deep-learning models to predict the radiotherapy outcome for BM. A novel self-attention-guided 3D residual network is introduced for predicting the outcome of local failure (LF) after radiotherapy using the baseline treatment-planning MRI. The 3D self-attention modules facilitate capturing long-range intra/inter slice dependencies which are often overlooked by convolution layers. The proposed model was compared to a vanilla 3D residual network and 3D residual network with CBAM attention in terms of performance in outcome prediction. A training recipe was adapted for the outcome prediction models during pretraining and training the down-stream task based on the recently proposed big transfer principles. A novel 3D visualization module was coupled with the model to demonstrate the impact of various intra/peri-lesion regions on volumetric multi-channel MRI upon the network's prediction. The proposed self-attention-guided 3D residual network outperforms the vanilla residual network and the residual network with CBAM attention in accuracy, F1-score, and AUC. The visualization results show the importance of peri-lesional characteristics on treatment-planning MRI in predicting local outcome after radiotherapy. This study demonstrates the potential of self-attention-guided deep-learning features derived from volumetric MRI in radiotherapy outcome prediction for BM. The insights obtained via the developed visualization module for individual lesions can possibly be applied during radiotherapy planning to decrease the chance of LF.

Predicting the outcome of radiotherapy in brain metastasis by integrating the clinical and MRI-based deep learning features

BACKGROUND

A considerable proportion of metastatic brain tumors progress locally despite stereotactic radiation treatment, and it can take months before such local progression is evident on follow-up imaging. Prediction of radiotherapy outcome in terms of tumor local failure is crucial for these patients and can facilitate treatment adjustments or allow for early salvage therapies.

PURPOSE

In this work, a novel deep learning architecture is introduced to predict the outcome of local control/failure in brain metastasis treated with stereotactic radiation therapy using treatment-planning magnetic resonance imaging (MRI) and standard clinical attributes.

METHODS

At the core of the proposed architecture is an InceptionResentV2 network to extract distinct features from each MRI slice for local outcome prediction. A recurrent or transformer network is integrated into the architecture to incorporate spatial dependencies between MRI slices into the predictive modeling. A visualization method based on prediction difference analysis is coupled with the deep learning model to illustrate how different regions of each lesion on MRI contribute to the model's prediction. The model was trained and optimized using the data acquired from 99 patients (116 lesions) and evaluated on an independent test set of 25 patients (40 lesions).

RESULTS

The results demonstrate the promising potential of the MRI deep learning features for outcome prediction, outperforming standard clinical variables. The prediction model with only clinical variables demonstrated an area under the receiver operating characteristic curve (AUC) of 0.68. The MRI deep learning models resulted in AUCs in the range of 0.72 to 0.83 depending on the mechanism to integrate information from MRI slices of each lesion. The best prediction performance (AUC = 0.86) was associated with the model that combined the MRI deep learning features with clinical variables and incorporated the inter-slice dependencies using a long short-term memory recurrent network. The visualization results highlighted the importance of tumor/lesion margins in local outcome prediction for brain metastasis.

CONCLUSIONS

The promising results of this study show the possibility of early prediction of radiotherapy outcome for brain metastasis via deep learning of MRI and clinical attributes at pre-treatment and encourage future studies on larger groups of patients treated with other radiotherapy modalities.

The radiomic-clinical model using the SHAP method for assessing the treatment response of whole-brain radiotherapy: a multicentric study

OBJECTIVE

To develop and validate a pretreatment magnetic resonance imaging (MRI)-based radiomic-clinical model to assess the treatment response of whole-brain radiotherapy (WBRT) by using SHapley Additive exPlanations (SHAP), which is derived from game theory, and can explain the output of different machine learning models.

METHODS

We retrospectively enrolled 228 patients with brain metastases from two medical centers (184 in the training cohort and 44 in the validation cohort). Treatment responses of patients were categorized as a non-responding group vs. a responding group according to the Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) criteria. For each tumor, 960 features were extracted from the MRI sequence. The least absolute shrinkage and selection operator (LASSO) was used for feature selection. A support vector machine (SVM) model incorporating clinical factors and radiomic features wase used to construct the radiomic-clinical model. SHAP method explained the SVM model by prioritizing the importance of features, in terms of assessment contribution.

RESULTS

Three radiomic features and three clinical factors were identified to build the model. Radiomic-clinical model yielded AUCs of 0.928 (95%CI 0.901-0.949) and 0.851 (95%CI 0.816-0.886) for assessing the treatment response in the training cohort and validation cohort, respectively. SHAP summary plot illustrated the feature's value affected the feature's impact attributed to model, and SHAP force plot showed the integration of features' impact attributed to individual response.

CONCLUSION

The radiomic-clinical model with the SHAP method can be useful for assessing the treatment response of WBRT and may assist clinicians in directing personalized WBRT strategies in an understandable manner.

KEY POINTS

• Radiomic-clinical model can be useful for assessing the treatment response of WBRT. • SHAP could explain and visualize radiomic-clinical machine learning model in a clinician-friendly way.

A priori prediction of local failure in brain metastasis after hypo-fractionated stereotactic radiotherapy using quantitative MRI and machine learning

This study investigated the effectiveness of pre-treatment quantitative MRI and clinical features along with machine learning techniques to predict local failure in patients with brain metastasis treated with hypo-fractionated stereotactic radiation therapy (SRT). The predictive models were developed using the data from 100 patients (141 lesions) and evaluated on an independent test set with data from 20 patients (30 lesions). Quantitative MRI radiomic features were derived from the treatment-planning contrast-enhanced T1w and T2-FLAIR images. A multi-phase feature reduction and selection procedure was applied to construct an optimal quantitative MRI biomarker for predicting therapy outcome. The performance of standard clinical features in therapy outcome prediction was evaluated using a similar procedure. Survival analyses were conducted to compare the long-term outcome of the two patient cohorts (local control/failure) identified based on prediction at pre-treatment, and standard clinical criteria at last patient follow-up after SRT. The developed quantitative MRI biomarker consists of four features with two features quantifying heterogeneity in the edema region, one feature characterizing intra-tumour heterogeneity, and one feature describing tumour morphology. The predictive models with the radiomic and clinical feature sets yielded an AUC of 0.87 and 0.62, respectively on the independent test set. Incorporating radiomic features into the clinical predictive model improved the AUC of the model by up to 16%, relatively. A statistically significant difference was observed in survival of the two patient cohorts identified at pre-treatment using the radiomics-based predictive model, and at post-treatment using the the RANO-BM criteria. Results of this study revealed a good potential for quantitative MRI radiomic features at pre-treatment in predicting local failure in relatively large brain metastases undergoing SRT, and is a step forward towards a precision oncology paradigm for brain metastasis.

Differentiation Between Glioblastoma Multiforme and Metastasis From the Lungs and Other Sites Using Combined Clinical/Routine MRI Radiomics

Background

Differentiation between cerebral glioblastoma multiforme (GBM) and solitary brain metastasis (MET) is important. The existing radiomic differentiation method ignores the clinical and routine magnetic resonance imaging (MRI) features.

Purpose

To differentiate between GBM and MET and between METs from the lungs (MET-lung) and other sites (MET-other) through clinical and routine MRI, and radiomics analyses.

Methods and Materials

A total of 350 patients were collected from two institutions, including 182 patients with GBM and 168 patients with MET, which were all proven by pathology. The ROI of the tumor was obtained on axial postcontrast MRI which was performed before operation. Seven radiomic feature selection methods and four classification algorithms constituted 28 classifiers in two classification strategies, with the best classifier serving as the final radiomics model. The clinical and combination models were constructed using the nomograms developed. The performance of the nomograms was evaluated in terms of calibration, discrimination, and clinical usefulness. Student’s t-test or the chi-square test was used to assess the differences in the clinical and radiological characteristics between the training and internal validation cohorts. Receiver operating characteristic curve analysis was performed to assess the performance of developed models with the area under the curve (AUC).

Results

The classifier fisher_decision tree (fisher_DT) showed the best performance (AUC: 0.696, 95% CI:0.608-0.783) for distinguishing between GBM and MET in internal validation cohorts; the classifier reliefF_random forest (reliefF_RF) showed the best performance (AUC: 0.759, 95% CI: 0.613-0.904) for distinguishing between MET-lung and MET-other in internal validation cohorts. The combination models incorporating the radiomics signature and clinical-radiological characteristics were superior to the clinical-radiological models in the two classification strategies (AUC: 0.764 for differentiation between GBM in internal validation cohorts and MET and 0.759 or differentiation between MET-lung and MET-other in internal validation cohorts). The nomograms showed satisfactory performance and calibration and were considered clinically useful, as revealed in the decision curve analysis.

Data Conclusion

The combination of radiomic and non-radiomic features is helpful for the differentiation among GBM, MET-lung, and MET-other.

Classification of true progression after radiotherapy of brain metastasis on MRI using artificial intelligence: a systematic review and meta-analysis

Background

Classification of true progression from nonprogression (eg, radiation-necrosis) after stereotactic radiotherapy/radiosurgery of brain metastasis is known to be a challenging diagnostic task on conventional magnetic resonance imaging (MRI). The scope and status of research using artificial intelligence (AI) on classifying true progression are yet unknown.

Methods

We performed a systematic literature search of MEDLINE and EMBASE databases to identify studies that investigated the performance of AI-assisted MRI in classifying true progression after stereotactic radiotherapy/radiosurgery of brain metastasis, published before November 11, 2020. Pooled sensitivity and specificity were calculated using bivariate random-effects modeling. Meta-regression was performed for the identification of factors contributing to the heterogeneity among the studies. We assessed the quality of the studies using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) criteria and a modified version of the radiomics quality score (RQS).

Results

Seven studies were included, with a total of 485 patients and 907 tumors. The pooled sensitivity and specificity were 77% (95% CI, 70–83%) and 74% (64–82%), respectively. All 7 studies used radiomics, and none used deep learning. Several covariates including the proportion of lung cancer as the primary site, MR field strength, and radiomics segmentation slice showed a statistically significant association with the heterogeneity. Study quality was overall favorable in terms of the QUADAS-2 criteria, but not in terms of the RQS.

Conclusion

The diagnostic performance of AI-assisted MRI seems yet inadequate to be used reliably in clinical practice. Future studies with improved methodologies and a larger training set are needed.

A Brain Tumor Segmentation Framework Based on Outlier Detection Using One-Class Support Vector Machine

Accurate segmentation of brain tumors is a challenging task and also a crucial step in diagnosis and treatment planning for cancer patients. Magnetic resonance imaging (MRI) is the standard imaging modality for detection, characterization, treatment planning and outcome evaluation of brain tumors. MRI scans are usually acquired at multiple sessions before and after the treatment. An automatic segmentation framework is highly desirable to segment brain tumors in MR images as it streamlines the image-guided radiation therapy workflow considerably. Automatic segmentation of brain tumors also facilitates an incremental development of data-driven systems for therapy outcome prediction based on radiomics analysis. In this study, an outlier-detection-based segmentation framework is proposed to delineate brain tumors in magnetic resonance (MR) images automatically. The proposed method considers the tumor and edema pixels in an MR image as outliers compared to the pixels associated with the healthy tissue. The framework generates two outlier masks using independent one-class support vector machines that operate on post-contrast T1-weighted (T1w) and T2-weighted-fluid-attenuation-inversion-recovery (T2-FLAIR) images. The outlier masks are subsequently refined and fused using a number of morphological and logical operators to estimate a tumor mask for each image slice. The framework was constructed and evaluated using the MRI data acquired from 35 and 5 patients with brain metastasis, respectively. The obtained results demonstrated an average Dice similarity coefficient and Hausdorff distance of 0.84 ± 0.06 and 1.85 ± 0.48 mm, respectively, between the manual (ground truth) and automatic tumor contours, on the independent test set.

A Cascaded Deep-Learning Framework for Segmentation of Metastatic Brain Tumors Before and After Stereotactic Radiation Therapy

Radiation therapy is a major treatment option for brain metastasis. For radiation treatment planning and outcome evaluation, magnetic resonance (MR) images are acquired before and at multiple sessions after the treatment. Accurate segmentation of brain tumors on MR images is crucial for treatment planning, response evaluation, and developing data-driven models for outcome prediction. Due to the high volume of imaging data acquired from each patient at multiple follow-up sessions, manual tumor segmentation is resource- and time-consuming in clinic, hence developing an automatic segmentation framework is highly desirable. In this work, we proposed a cascaded 2D-3D Unet framework to segment brain tumors automatically on contrast-enhanced T1- weighted images acquired before and at multiple scan sessions after radiotherapy. 2D Unet is a well-known structure for medical image segmentation. 3D Unet is an extension of 2D Unet with a volumetric input image to provide richer spatial information. The limitation of 3D Unet is that it is memory consuming and cannot process large volumetric images. To address this limitation, a large volumetric input of 3D Unet is often patched to smaller volumes which leads to loss of context. To overcome this problem, we proposed using two cascaded 2D Unets to crop the input volume around the tumor area and reduce the input size of the 3D Unet, obviating the need to patch the input images. The framework was trained using images acquired from 96 patients before radiation therapy and tested using images acquired from 10 patients before and at four follow-up scans after radiotherapy. The segmentation results for the images of independent test set demonstrated that the cascaded framework outperformed the 2D and 3D Unets alone, with an average Dice score of 0.9 versus 0.86 and 0.88 for the baseline, and 0.87 versus 0.83 and 0.84 for the first followup. Similar results were obtained for the other follow-up scans.