Radiomic-based solutions
The term “radiomics” first appeared in literature in 2012 through an article published by Lambin et al. (22). This approach, focused on medical imaging data, aims to extract a large set of features from an image for a better characterization of tumor. Radiomic protocols require the following six steps: image acquisition, image reconstruction and pre-processing, segmentation, resampling, features extraction, and features selection and model-based feature construction (23). Because of these various steps, the use of radiomics aims to be potentially predictive compared to imaging biomarker analyses based on basic features such as mean or peak intensity. Image characteristics are subjected to a more in-depth analysis, making the features more relevant for prediction, and consequently, the results are more effective. Radiomics models are capable of predicting therapeutic response or overall survival (23).
In that context, the use of radiomics to develop models capable of predicting treatment response prior to brain tumor treatment initiation has been explored in several studies.
One of these studies (24) investigated the extraction of radiomic features from post-treatment MRI in patients with BM to predict local tumor control with an estimation of the tumor volume percentage compared to pre-treatment and overall survival with 256 and 237 patients, respectively. Three models were constructed through the training of SVMs using a Gaussian kernel and Bayesian optimization for hyperparameter tuning: (i) clinical features (age, gender overall survival, numbers of tumors, local tumor control, and median dose), (ii) radiomic features, and (iii) combined clinical and radiomic features. For both prediction objectives, the model combining clinical and radiomic features achieved very interesting performances with an area under the receiver operating characteristic curve (AUC) of 0.95 for local tumor control and 0.82 for overall survival.
Furthermore, a clinical study (25) was conducted to predict survival stratification of 125 patients with GBM. Radiomic features were extracted from MRI images. Among the three tested ML models, the SVM model demonstrated the best performance, with an AUC of 0.92.
Table 1B
(24–28) summarizes several studies on the prediction of treatment response based on radiomics obtained from pre-treatment imaging. In all studies, the AUC is between 0.62 and 0.95. All these studies highlighting combining radiomic features with clinical features enhance prediction performance. However, radiomics has some limitations for routine clinical application. Most published studies have a relatively small patient cohort especially for GBM. However, to develop effective models, a sufficiently large training and test set is mandatory (29). Because of its complexity, radiomics presents the challenge of low interpretability of the features and models used, raising caution among physicians regarding the use of radiomics models in clinical settings (30). Beyond these points, the main limitation of radiomics remains the low stability and inter-hospital portability of the models (29). To resolve to this challenge, initiatives like the “Imaging Biomarker Standardization Initiative (IBSI)” (31) have been developed to complement radiomic features extraction; however, the robustness of these predictive models remains an issue before their adoption as a standard of care as shown by Peerlings and colleagues (32) for diffusion MRI or CT (33) or even for Test–Retest in PET imaging (34).
Therefore, it is timely to explore more innovative current developments in AI that may enable predictive characterization of treatment efficacy. DL is known to be able to extract more complex and a larger number of features in medical imaging than radiomics, which could lead to better performance (35).
Artificial intelligence-based solutions
Several studies have evaluated the use of AI algorithms to assess the therapeutic efficacy of GBM and BM. A clinical study (36) involving 124 patients with BM developed a CNN-based architecture to extract features from each MRI slice to predict the outcome of local control/failure in BM treated with stereotactic radiation therapy. A CNN is a type of DL neural network specifically designed to process structured data arrays, such as images. They integrated an InceptionResentV2 CNN architecture and a transformer (to consider spatial dependences between MRI slices during modeling). Depending on the mechanism of integration of information from each MRI slice, the AUC ranged from 0.72 to 0.86. The best performance was obtained with the combination of DL features obtained from anatomical MRI with clinical variables (tumor size, age, gender, tumor location, histology, total dose, previous WBRT, and number of BM).
In a study including 30 patients (15 with low-grade glioma and 15 with GBM), Vollmuth et al. (37) demonstrated that AI using artificial neural network (ANN) for brain and then tumor segmentations has the potential to provide a more reproducible and standardized assessment of treatment response on MRI compared to manual two-dimensional measurements of tumor burden using RANO criteria. Time to progression (TTP) was initially evaluated according to RANO criteria based on MRI and then revaluated by incorporating additional information from AI-enhanced MRI sequences that describe longitudinal changes in tumor volume. The inter-observer concordance correlation coefficient (CCC) for TTP measurements was 0.77 using the RANO criteria alone. With the addition of AI, the CCC increased to 0.91. This improvement was most observed in patients with low-grade gliomas (0.70 without AI vs. 0.90 with AI). Because of the less aggressive nature of these tumors, reliable assessment of TTP can be more difficult.
In a previous study, Luckett et al. (38) show a good performance with an accuracy of 90.6% in classifying survival (<1 year, 1–2 years, and >2 years), employing a deep feedforward CNN comprising three hidden layers with eight neurons in each layer to predict patient survival in a cohort of 133 individuals. Ortega-Martorell and colleagues also showed a good performance of one-dimension CNN in a preclinical study to track therapy response in GBM (39). The 1D-CNN performed better than different ML models, showing the superiority of DL methods.
Our review of the literature reveals that the CNN exhibits superior performance. Although the architecture is not novel, it is particularly suited to medical imaging and currently offers the most effective means of predicting treatment efficacy (40).
Table 1C
(36–39) summarizes several studies on the prediction of treatment response based on AI algorithms from pre-treatment MRI. As in the radiomics-based studies, the best performance is achieved by combining imaging data with clinical information.
Many studies applying AI in this field are based on relatively small data cohorts (less than 100 for GBM). However, a large data cohort is essential for optimal training of AI models (41). Centralizing a large amount of data in a single center can be challenging, and the performances of models are not always transferable between centers. Federated learning (42) addresses this issue by enabling learning from distributed data without transferring it between sites. Federated learning is a DL paradigm in which a model is trained across multiple decentralized devices or servers located in various medical centers, each holding local data samples, without the need to exchange the raw data. The only parameters shared among the different hospitals are the model parameters, not the raw medical data.
In addition, AI methodology is constantly evolving and new architectures appear every year. The models we have presented in this review give an overview of what is being done today, but new architectures such as diffusion models or full transformers should be more present in the years to come. One example is the UNEt TRansformer (UNETR) (43), which adapts the CNN encoder/decoder models proposed by UNET to transformer architectures in order to process sequential representations of the input volume more efficiently. Transformers are a type of AI model designed to efficiently process sequential data, such as text. Functional imaging such as proliferation index or other indicators is more relevant for assessing therapeutic efficacy (17). To our knowledge, no study involving AI models uses functional imaging biomarkers for predicting GBM efficacy as all the articles reported in this review used clinical routine anatomical MRI. However, in other cancers with radiomic models, Knuth and colleagues as well as Zhang and colleagues support the add value of function biomarkers in comparison to anatomical MRI in rectal (44) and breast cancers, respectively (45).
Opting for more functional imaging biomarkers instead of anatomical MRI could potentially improve AI performance in predicting treatment efficacy.
It is important to note that current studies were based on 2016 WHO classification rather than the 2021 one. To the best of our knowledge, no study has yet evaluated the potential of AI models to predict treatment outcomes of GBM according to the WHO 2021 classification. These models may not fully reflect current standards and advancements in the field, potentially leading to biases in predictions. However, current performances of the AI models to predict treatment outcome are still valid if they do not take into account the grade of the tumor, for example, if the input data only take the pre-treatment MRI. If the model is capable of predicting the treatment outcome of a brain lesion on an MRI, it should still be able to do so regardless of whether the brain lesion is designated as a GBM or a grade 4 astrocytoma Therefore, it is essential to incorporate recent classifications to ensure that AI models are aligned with best clinical practices and provide reliable and relevant recommendations.