Using ML to predict histologic subtype and grade of canine gliomas

With radiomics, MRI textural analysis is correlated with glioma subtype and grade

Machine learning predicts histologic type and grade of canine gliomas based on MRI texture analysis

Pablo Barge, Anna Oevermann, Arianna Maiolini, Alexane Durand


Texture Analysis (TA) mathematically extracts quantitative information from medical images and provides objective measurements of tumor heterogeneity. Texture features can be used as input for machine learning (ML) models, which have shown promise in differentiating various pathologies in small animal medicine. However, previous attempts to use TA-based ML classifiers for canine gliomas have been unsuccessful. The study aims to overcome these limitations by using multiple segments across the entire tumor volume, which has not been done in canine patients before. The researchers hypothesize that this approach, combined with different ML classifiers, will lead to a more accurate classification of glioma types and grades based on MRI-TA compared to conventional MRI.

Materials and Methods

In this study, the researchers used neuroimaging techniques to analyze brain tumors. The study design was presented in a figure, and images were converted into a specific format. Different types of brain images were aligned using software, and the tumor's enhancing and non-enhancing parts, as well as the perilesional vasogenic edema, were manually segmented by two experts. The enhancing segment referred to areas of the tumor that showed contrast enhancement, while the non-enhancing segment represented tumor areas without enhancement. The vasogenic edema segment included hyperintense areas in the peri-tumoral white matter. Certain regions, such as meninges, large vessels, intratumoral cysts with suppressing FLAIR signals, and areas affected by partial volume averaging artifact, were not included in any of the segments. The segmentations were then exported as label-maps to extract texture features for further analysis.

Pre-processing techniques and texture feature extraction were carried out using a freeware software called LIFEx. The goal of pre-processing was to standardize the images in terms of pixel spacing, gray-level intensities, and gray-level histogram bins. Image interpolation was done to ensure a consistent in-plane resolution of 1x1 mm while preserving the original slice thickness. Image discretization involved using a fixed bin number of 32, and the gray levels of the MRI within the segmentations were normalized to the mean plus or minus three standard deviations. A total of 61 texture features were extracted per segment in each sequence, resulting in a maximum of 732 texture values per tumor.

Supervised ML algorithms such as Support Vector Machine (SVM), Random Forest (RF), and k-nearest neighbors (kNN) were used for classification and regression problems in medical imaging. The models were programmed using Python and various libraries and run on Google Colab. Before feeding the models, the data was normalized using z-score normalization and dimensionality reduction was performed using Principal Component Analysis (PCA). A leave-one-out cross-validation technique was used to assess the performance of the models. Different datasets were prepared, including all extracted texture features and smaller datasets with specific segments and sequences. The models aimed to predict glioma histological type and grade, and the predictions were compared to histopathological diagnosis to calculate accuracy, sensitivity, specificity, and AUC. The study included a total of 13 datasets and excluded oligosarcomas from the multiclass models due to the low number of cases.


Machine learning (ML) classifiers were used to predict different types of tumors in brain imaging data. The multiclass ML classifiers had an average accuracy of 77% for predicting tumor types, with varying sensitivity and specificity values for different tumor types. The binary ML classifiers had an average accuracy of 76% for predicting high-grade gliomas, with better performance from SVM and kNN classifiers than RF. The ML models were not fed with the enhancing segment as only a few cases showed enhancement. As a result, further analysis was conducted on a smaller dataset. The SVM multiclass classifier performed well for predicting tumor types on edema segments, while the SVM and kNN binary classifiers achieved the best performance for predicting high-grade gliomas on non-enhancing segments in T2w and T1w postcontrast sequences. Overall, the classifiers had varied performances on different sequences and segments for predicting tumor types and high-grade gliomas.


The researchers found that TA-based ML models using multiple segments of the tumor volume provided more accurate classification compared to conventional MRI. The SVM and kNN models demonstrated higher accuracy, sensitivity, and specificity in distinguishing between high-grade and low-grade gliomas compared to conventional MRI. The study also highlighted the potential of TA-based ML in improving treatment planning and prognosis assessment for canine gliomas. However, there were limitations such as the imbalanced dataset and difficulty in differentiating between glioma types. The study suggests the need for further research to explore the different texture metrics across glioma types and grades. The use of automated techniques for tumor segmentation in dogs is also recommended for future studies.


The authors conclude that machine learning models using MRI-TA can effectively distinguish different types and grades of intracranial canine gliomas. They found that the support vector machine (SVM) model had accuracies of up to 94% for tumor type discrimination and 87% for tumor grade discrimination. The texture features that were most useful for differentiating tumor types were associated with peri-tumoral edema in T1w images, while those for tumor grades were related to the non-enhancing part of the tumor in T2w images. However, the authors recommend conducting further multicenter studies with larger sample sizes to validate and strengthen these findings.

This article was summarized by an AI tool that uses natural language processing. The tool is not perfect and may make mistakes or produce inaccurate or irrelevant information, but is reviewed by the post’s author prior to publishing. If you want to learn more about the article, please refer to the original source that is cited at the end of the article.