A Study on Classification and Segmentation of Brain Tumor MRI using MediAI

 A Study on Classification and Segmentation of Brain Tumor MRI using MediAI

http://dx.doi.org/10.4283/JMAG.2025.30.1.74

Abstract

Brain tumors, arising due to genetic, environmental, and immune system factors, are typically classified as primary or metastatic. Primary brain tumors originate from brain tissue and include types such as adenomas, epithelial tumors, gliomas, meningiomas, and schwannomas, while metastatic tumors spread from other body parts. The rapid and accurate diagnosis of brain tumors is crucial, and MRI is an indispensable tool in this regard due to its non-invasive nature and superior imaging capabilities. This study focuses on developing a method for the classification and segmentation of MRI images of primary brain tumors, particularly gliomas, to enhance diagnosis and treatment assessment. We propose MediAI, a Convolutional Neural Network (CNN) leveraging transfer learning with ResNet50, to classify MRI images of brain tumors. The dataset comprises 414 MRI images of primary brain tumors (86 adenomas, 84 epithelial tumors, 82 gliomas, 80 meningiomas, and 82 schwannomas) and 39 normal MR images, sourced from various public datasets. Experimental results demonstrated that MediAI achieved an impressive classification accuracy of 97.6 %. For tumor segmentation, we applied anisotropic diffusion filtering, followed by thresholding using the Otsu method, to accurately detect and delineate tumor regions, particularly in glioblastomas, highly malignant brain tumors. Tumor regions were further refined through morphological operations, and the final tumor contours were extracted and overlaid on the original images. Performance evaluation of MediAI was conducted using a confusion matrix, calculating precision, recall, and F1-score for each tumor type. Results indicated that gliomas, in particular, were classified with a precision of 91 %,a recall of 99 %, and an F1-score of 95 %. A comparison with existing studies demonstrated that MediAI outperforms previous methods, achieving the highest reported accuracy for brain tumor classification at 97.6 %. The proposed methodology not only facilitates accurate brain tumor classification but also enhances the monitoring of treatment responses by tracking changes in segmented tumor regions. Future work will focus on utilizing Radiomics to map tumor, necrotic, and edema regions for advancing diagnostic and therapeutic paradigms in glioma treatment. 

1. Introduction 

Brain tumors develop due to genetic predispositions, environmental factors, and immune system abnormalities. Specific genetic mutations can disrupt the regulation of cellular growth, potentially leading to tumor formation. Additionally, abnormalities in intracellular signaling pathways can influence cellular proliferation and survival, contributing to tumorigenesis [1]. Chronic inflammation plays a critical role in tumor development by promoting cell growth and survival. Certain hormones also affect cellular growth, with some brain tumors linked to hormonal changes. Brain tumors are classified into primary and metastatic types [2]. Primary brain tumors originate from specific types of brain cells or tissues within the brain, including adenomas, epithelial tumors, glioblastomas, meningiomas, and schwannomas. In contrast, metastatic brain tumors result from the spread of cancer cells from other body parts, such as lung, breast, or skin cancer, which disseminate to the brain via the bloodstream. Metastatic tumors can spread from their original site, leading to cancer manifestation in multiple locations [3]. The diagnosis of brain tumors typically involves neurological exams, biopsies for histological analysis, blood tests to measure specific tumor markers, and imaging techniques like MRI, CT, and PET scans. 

Among these modalities, MRI has proven to be an indispensable tool for the rapid and accurate diagnosis of brain tumors due to its non-invasive nature and versatile imaging capabilities [4]. Glioblastoma, a form of glioma, originates from astrocytes in the brain and is characterized by high malignancy, making it the most common primary intracranial tumor in the central nervous system [5]. While glioblastomas are most frequently located in the cerebral hemispheres, they can also occur in other brain or spinal cord regions. The prognosis for glioblastoma is poor, with survival rates averaging around five years [6]. On MRI, glioblastomas appear as regions of high signal intensity with surrounding edema on T2-weighted and FLAIR images, while they show low signal intensity on T1-weighted images. Contrast-enhanced imaging reveals significant enhancement due to the disruption of the blood-brain barrier (BBB) [7]. As highly aggressive tumors with a poor prognosis, glioblastomas rely heavily on MRI for diagnosis and assessment. Diffusion-weighted imaging (DWI) is especially valuable for evaluating glioblastomas, as it is sensitive to changes in the movement of water molecules within the tissue. On DWI, the high cellular density of glioblastomas restricts diffusion, resulting in increased signal intensity. Apparent diffusion coefficient (ADC) maps, derived from DWI, quantify diffusion. Due to their high cellular density, glioblastomas typically exhibit low ADC values [8]. However, necrotic regions within the tumor may contain fluid, which can lead to elevated ADC values in these areas. DWI can also reflect changes in cellular density during treatment: decreased cellular density increases diffusion, leading to higher ADC values, while increased cellular density restricts diffusion, resulting in lower ADC values. These ADC values can serve as biomarkers for evaluating treatment response in glioblastomas, with newly restricted diffusion indicating tumor progression or recurrence [9]

2. Materials and Methods 

2.1. Brain Tumor MRI Dataset 

The data set for the study consisted of 414 primary brain tumor MR images of 86 adenomas, 84 epithelial tumors, 82 glioblastomas, 80 meningiomas, 82 schwannomas, and 39 normal MR images, for a total of 453 images. The images were collected from GitHub and NEJM [10], Auntminnie [11], Medscape [12], Radiology cases [13], and radiopaedia [14] sites, labeled by brain tumor, and organized into a file named ‘BrainTumors_1.1.zip’. 

Fig. 1. Part of the MR images of brain tumors for the experiment (Adenoma, Epithelioma, Glioblastoma, Meningioma, Schwannoma).

Fig. 1 shows a portion of the MR images of Adenoma, Epithelioma, Glioblastoma, Meningioma, and Schwannoma brain tumors constructed for the experiment. 

2.2. Brain Tumor Classification using CNN 

The Convolutional Neural Network (CNN) developed for brain tumor classification, referred to as MediAI, was constructed using a transfer learning approach based on ResNet50. 

Fig. 2. (Color online) Structure of MediAI that classifies brain tumors using ResNet transfer learning

Similar to most CNN architectures, MediAI aims to improve recognition accuracy and reduce training time by utilizing feature values extracted from input images as training data. This is achieved through a repetitive process of convolution and pooling, where feature extraction and down-sampling occur iteratively. During this iterative process, the characteristics of the input images are progressively refined and abstracted with each successive layer [15]. In MediAI, the flattening stage transforms the feature maps generated by the convolutional and pooling layers into a one-dimensional vector. By concatenating multiple independent one-dimensional vectors within a dense layer—specifically, a fully connected layer—a multilayer perceptron (MLP) is formed [16]. The MLP classifies the images through the application of activation functions. For hyperparameter optimization in MediAI, Bayesian Optimization was employed. This method selects the next hyperparameter combination based on previous results. A model was trained for each combination of hyperparameters, and performance was evaluated using a validation dataset. Precision, recall, and F1-score were used as the performance metrics [17]. Fig. 2 illustrates the structure of MediAI, which classifies brain tumors using ResNet-based transfer learning. 

2.3. Tumor Region Segmentation

 Following the classification of brain tumors in MRI scans, tumor region detection and segmentation are carried out on selected specific brain tumor MRIs according to the steps outlined in Table 1. 

In the image loading phase, upon selecting JPEG, BMP, or PNG files, the corresponding image paths are retrieved, and the images are loaded accordingly. The imshow() function is used to display the loaded images on the screen. During the filtering stage, Anisotropic Diffusion is applied to reduce noise, after which the filtered images are resized to 256 × 256 pixels. If the images are in color, they are then converted to grayscale. Segmentation is performed by calculating the optimal threshold using the Otsu method on the converted image. The Otsu method is an algorithm that automatically determines the optimal threshold to separate an image into two classes (background and object) based on the image's histogram. This method minimizes the variance within each class while maximizing the distance between the two classes. In other words, the optimal threshold is set at the point where the inter-class variance is maximized [18]. During the thresholding process, the images are binarized to distinguish between the tumor regions and the background. The morphological operations phase involves labeling the binary images and calculating the characteristics of each object to estimate the tumor. In the bounding box drawing stage, bounding boxes are drawn around the identified tumor regions. The tumor contour extraction phase involves applying erosion operations to extract the tumor contours, and in the final image overlay stage, the tumor contours are highlighted in red on the original images. 

3. Experiment and Result 

The experiments for segmenting tumor regions in glioblastomas from brain MRI were conducted following the procedure illustrated in Fig. 3. 

Fig. 3. (Color online) Experimental procedures for brain tumor classification and glioblastoma segmentation

Using the MediAI CNN, the compressed images in the dataset were extracted and stored in the designated folder. The images input into MediAI underwent feature extraction to enhance training efficiency and improve recognition accuracy, with the extracted feature values serving as the input data. This process involved iterative applications of convolution and pooling operations, which progressively refined and abstracted the feature values. During the flattening stage, the feature maps were transformed into a one-dimensional vector, which was then classified using activation functions. Fig. 4 shows brain tumor images classified by the MediAI training process. 

Fig. 4. Brain tumor images classified by MediAI training

To assess the accuracy of the classified MediAI images, a confusion matrix was created, as shown in Fig. 5. 

Fig. 5. (Color online) MediAI's confusion matrix for classification of brain tumor MR images.

Additionally, a receiver operating characteristic (ROC) curve was plotted to evaluate the performance of MediAI, as presented in Fig. 6. 

Fig. 6. (Color online) MediAI's ROC Curve for Classifying Brain Tumor MR Images.

As illustrated in Fig. 6, the accuracy was found to be 97 %. To facilitate the segmentation of tumor regions in brain tumors classified by MediAI, preprocessing was performed through filtering to reduce noise, and the images were resized to 256 × 256 pixels. The filtered images are shown in Table 2. 

Among the filtered images, those predicted to be glioblastoma were selected, as shown in Fig. 7. 

Fig. 7. Images classified as glioblastoma in MediAI's brain tumor MRI classification experiment.

The selected image underwent a thresholding process to binarize the image and separate the tumor regions from the background. During this process, the threshold was automatically adjusted to accurately detect the tumor regions. Based on the identified tumor regions, the tumor segmentation stage involved marking the tumor area with a red bounding box. 

Fig. 8. (Color online) Bounding box image in a brain glioblastoma segmentation experiment. 

Fig. 8 presents an image with a bounding box in a glioblastoma segmentation experiment. Subsequently, the extracted tumor contours are presented in Fig. 9. 

Fig. 9. (Color online) Tumor outline obtained by performing erosion.

Finally, the tumor contours were overlaid onto the original images. Fig. 10 shows the images with the tumor contours overlaid. 

Fig. 10. (Color online) Detection of glioblastoma.

4. Discussion 

This study proposes a method for classifying MRI scans of primary brain tumors, including adenomas, epithelial tumors, gliomas, meningiomas, and schwannomas, by disease type, with a particular focus on segmenting the tumor regions of gliomas. To improve the data dependency problem and interpretability of MediAI and to verify whether the generalized prediction for various data sets is reliable, we conducted an experiment using data augmentation techniques as shown in Fig. 11. 

Fig. 11. (Color online) Image processing-based data augmentation.

The results of the experiment showed that the difference between the original data set and the augmented data set was as shown in Fig. 12. 

Fig. 12. (Color online) Experimental results (AUC) before and after data augmentation.

The reason why the difference before and after data augmentation is not large is thought to be because MediAI, a transfer learning technique, uses dropout to solve the data overfitting problem. Dropout is a method of proceeding with learning while omitting some neurons in the fully connected layer. It changes the values of some neurons to 0 so that they have no effect on the forward pass and backpropagation. Dropout is applied during training, and all neurons are used during testing. ResNet50 outperforms previous AlexNet in medical imaging research due to its deep residual learning function, high accuracy, efficiency, scalability, and overfitting problem-solving by skin connection. 

Fig. 13. (Color online) Performance evaluation results of ResNet50 and AlrexNet after data augmentation.

The experimental results to compare the performance of Resnet50 and AlrexNet are shown in Fig. 13. Brain tumor classification in the proposed MR images was performed using MediAI, achieving an accuracy of 97.6 %. Additionally, the precision, recall, and F1-score for the classified Brain Tumor types, calculated using a confusion matrix, are presented in Fig. 14. 

Fig. 14. (Color online) Precision, recall, and F1-score for Brain Tumor types calculated using a confusion matrix.

To validate the effectiveness of this study, previous research analyzing disease classification in MRI scans was reviewed. In the study "Convolutional Neural Networks for MultiClass Brain Disease Detection Using MRI Images" by Muhammed Talo et al., pre-trained models such as AlexNet, VGG-16, ResNet-18, ResNet-34, and ResNet50 were used to automatically classify MRI scans into categories of normal, cerebrovascular diseases, neoplasms, degenerative, and inflammatory diseases. Among these five pre-trained models, ResNet-50 achieved the highest classification accuracy of 95.23 % ± 0.6 % [19]. In the research titled "Brain Tumor Detection in Brain MRI Using Deep Learning" by Abhishek Anil et al., AlexNet achieved an accuracy of 89.64 %, VGG16 reached 93.22 %, and VGG19 achieved 95.78 % accuracy [20]. Driss Lamrani et al. reported in their study "Brain Tumor Detection Using MRI Images and Convolutional Neural Networks" that pre-trained architecture models achieved 96 % precision and classification accuracy [21]. Soheila Saeedi et al. proposed a 2D CNN in "MRIBased Brain Tumor Detection Using Convolutional Deep Learning Methods and Selected Machine Learning Techniques," achieving an accuracy of 96.47 %, with the learning accuracy of the proposed autoencoder network being 95.63 %. P. Gokila Brindha et al. mentioned that the CNN model applied to the test data in their study "Brain Tumor Detection in MRI Images Using Deep Learning Techniques" achieved an accuracy of 89 % [22, 23]. Finally, Javeria Amin et al. reported an average accuracy of 97.1 % in their research "A Unique Approach to Brain Tumor Detection and Classification Using MRI," as shown in Fig. 15. 

 

Fig. 15. (Color online) Comparison of accuracy of MediAI and other experiments in disease classification in brain tumor MRI

As seen in this study, the implemented MediAI achieved the highest accuracy of 97.6 % [24]. Glioblastoma is a highly aggressive brain tumor with a poor prognosis, and MRI plays a crucial role in its diagnosis and evaluation. Within MRI, DWI is particularly useful for assessing glioblastomas, as it is sensitive to changes in the movement of water molecules within the tissue. This study focuses on classifying brain tumors in MRI and, in particular, on detecting and segmenting the tumor regions in highly aggressive glioblastoma MRI DWI scans. Anisotropic diffusion filtering was initially applied to the selected glioblastoma MRI DWI scans to continuously filter the images and reduce contrast between pixels. The images were then resized, and thresholding techniques were used to convert the images into binary form. This filtering process aimed to separate potential tumor locations. Subsequently, morphological operations were applied to the pre-processed images to extract robust information from the potential tumor regions. Through this methodology, tumor regions were delineated based on the statistical average of glioblastoma involvement in MRI DWI. While this approach provided accurate results in most cases, difficulties arose when the tumor size was too small or when the tumor exhibited heterogeneous distribution patterns. 

5. Conclusion 

This study presents a refined approach to classifying MRI scans related to primary brain tumors, including adenomas, epithelial tumors, gliomas, meningiomas, and schwannomas, with a particular emphasis on the segmentation of glioma tumor regions. To achieve this classification, MediAI was meticulously developed using a transfer learning methodology based on a convolutional neural network, ResNet50. The dataset used for this study was constructed from 414 primary brain tumor MRI images, including 86 adenomas, 84 epithelial tumors, 82 gliomas, 80 meningiomas, and 82 schwannomas, as well as 39 normal MR images. These images were sourced from GitHub, NEJM, AuntMinnie, Medscape, Radiology Cases, and Radiopaedia, and compiled into the file "BrainTumors_1.1.zip." Experimental results showed that brain tumor MR images were classified with an accuracy of 97.6 %. To rigorously evaluate MediAI’s performance, a confusion matrix was employed to calculate precision, recall, and F1-score for each tumor category. • The results revealed that adenomas achieved a precision of 76.0 %, a recall of 84.0 %, and an F1-score of 80 %. • Epithelial tumors exhibited a precision of 99.0 %, recall of 94.0 %, and an F1-Score of 96 %. • Gliomas achieved a precision of 91.0 %, recall of 99.0 %, and an F1-Score of 95 %. • Meningiomas showed a precision of 92.0 %, recall of 89.0 %, and an F1-Score of 90 %. • In contrast, schwannomas achieved a precision of 98.0 %, recall of 95.0 %, and an F1-Score of 96 %. A comparative analysis with existing literature underscores the superior accuracy of this study, with MediAI achieving a 97.6 % accuracy. Within the spectrum of brain tumors, gliomas are known for their aggressive nature and poor prognosis. This study performed a series of experiments specifically aimed at detecting and segmenting tumor regions within glioma images. The results confirm the accurate detection and segmentation of brain tumors, facilitating the differentiation of tumor regions within the identified areas. This capability not only enhances the accuracy of glioma diagnosis, the deadliest form of brain tumor, but also enables monitoring the changes in segmented tumor regions throughout the treatment process, thereby aiding in the evaluation of therapeutic efficacy. Future research will focus on advancing diagnostic and assessment paradigms for glioma treatment efficacy by mapping the identified tumor, necrotic, and edema regions using Radiomics methodologies.

References 

[1] R. N. Curry and S. M. Glasgow, Front Cell Dev. Biol. 9, 659055 (2021). 

[2] S. Hibino, T. Kawazoe, H. Kasahara, S. Itoh, T. Ishimoto, M. S. Yanagimoto, and K. Taniguchi, Int. J. Mol. Sci. 22, 5421 (2021). 

[3] J. Ricardo, M. Figueroa, and E. Q. Lee, Am J. Med. 131, 874 (2018). 

[4] W. B. Overcast, K. M. Davis, C. Y. Ho, G. D. Hutchins, M. A. Green, B. D. Graner, and M. C. Veronesi, Curr. Oncol. Rep. 23, 34 (2021). 

[5] A. A. Abd-Elghany, A. A. Naji, B. A, H. Aldosary, M. A. Alsufayan, M. Alnasser, A. M. Ebtsam, and M. Z. Mahmoud, J. Radiat. Res. Appl. Sci. 12, 289 (2019). 

[6] A. K. Altwairgi, W. Algareeb, G. Yahya, and A. M. Maklad, Mol. Clin. Oncol. 4, 756 (2016). 

[7] K. Y. Shim, S. W. Chung, J. H. Jeong, I. Hwang, C. K. Park, T. M. Kim, S. H. Park, J. K. Won, J. H. Lee, S. T. Lee, R. E. Yoo, K. M. Kang, T. J. Yun, J. H. Kim, C. H. Sohn, K. S. Choi, and S. H. Choi, Sci. Rep. 11, 9974 (2021). 

[8] H. Lee, H. O. Kim, and Y. M. Huh, J. Magn. 25, 567 (2020). 

[9] A. Afaq, A. Andreou, and D. M. Koh, Cancer Imaging. 10, 179 (2010). 

[10]https://github.com/SartajBhuvaji/Brain-Tumor-Classification-DataSet/tree/master/Testing 

[11] https://my.auntminnie.com/cases/ 

[12] https://search.medscape.com/search?q=Brain%20Tumor %20MRI&plr=ref 

[13] https://mrionline.com/case-of-the-week 

[14] https://radiopaedia.org/search?q=Brain+MRI&scope=all &lang=us 

[15] R. Yamashita, M. Nishio, Do, R. K. G. Do, and K. Togashi, Insights into Imaging 9, 611 (2018). 

[16] X. Zhao, L. Wang, Y. Zhang, X. Han, M. Deveci, and M. Parmar, Artif. Intell. Rev. 57, 10462 (2024). 

[17] J. Snoek, H. Larochelle, and R. P. Adams, Advances in Neural Information Processing Systems. 25, 2951 (2012). 

[18] N. Otsu, IEEE Transactions on Systems, Man, and Cybernetics. 9, 62 (1979). 

[19] M. Talo, O. Yildirim, U. B. Baloglu, G. Aydin, and U. R. Acharya, Comput. Med. Imaging Graph. 78, 312 (2019). 

[20] A. Anil, A. Raj, H. A. Sarma, and N. C. R. Deepa, International Journal of Innovative Research in Applied Sciences and Engineering. 3, 15 (2019). 

[21] D. Lamrani, B. Cherradi, O. El Gannour, M. A. Bouqentar, and L. Bahatti, Int. J. Adv. Comput. Sci. Appl. 13, 452 (2022). 

[22] S. Saeedi, S. Rezayi, H. Keshavarz, and S. R. N. Kalhori, BMC Med. Inform. Decis. Mak. 23, 2144 (2023). 

[23] P. G. Brindha, M. Kavinraj, P. Manivasakam, and P. Prasanth, IOP Conf. Ser.: Mater. Sci. Eng. 1055 (2021). 

[24] J. Amin, M. Sharif, M. Yasmin, and S. L. Fernandes, Pattern Recognit. Lett. 139, 118 (2020).

by Ph. D., M. D. Hwunjae Lee