Cat swarm optimized tumor segmentation and an ensemble predictive model for glioblastoma patient sur
* Author to whom correspondence should be addressed.
Sigma Journal of Engineering and Natural Sciences 2025, Vol. 43, Issue 5, pp. 1592-1606; doi.org/10.14744/sigma.2025.00029
Abstract
Keywords: Cat Swarn Optimization; Glioblastoma; Stereotactic Radio surgery; Survival Prediction; Deep Learning; ResNet; U-Net; Ensemble Prediction; Random Forest; Tumor Segmentation
Introduction
The High degree of recurrence, poor prognosis, tumor heterogeneity, and limited treatment options are the main
issues in Glioblastoma. Glioblastomas’ complex and heterogeneous nature contribute to poor overall survival and also to treatment issues. While we see an increase in the use of Stereotactic Radiosurgery (SRS) which is very precise in
*Corresponding author. *E-mail address: bhagyareddy20380@gmail.com, bk8019@srmist.edu.in This paper was recommended for publication in revised form by Regional Editor Ahmet Selim Dalkilic Published by Yıldız Technical University Press, İstanbul, Turkey © Author. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
its delivery of radiation therapy to the tumor thus sparing the surrounding healthy tissue we are still have issues in accurate patient outcome prediction. In the area of artificial intelligence (AI) which includes deep learning (DL) we are seeing very promising results in the identification and management of glioblastomas’ but still accurate prognosis is impaired by tumor complexity. Our research has put forth a new approach which is a combination of Cat Swarm Optimization-assisted hybrid ResNet and U-Net architectures for accurate tumor segmentation with machine learning algorithms for very robust survival prediction. This we have done to out perform traditional models which we do so by improving segmentation accuracy and at the same time we are able to present more reliable survival predictions. We are into this with the aim to improve individualized treatment plans for GBM patients that are undergoing stereotactic radiosurgery. What we have done not only improves the care we give to glioblastoma patients but also we put forth the wide scale promise of AI driven methods in oncology and personal medicine. Our work gives out that which may be applied across many cancer types which in turn bridges the gap between very advanced AI techniques and clinical practice.
Results
dice score for the tumor core (TC), whole tumor (WT), and enhancing tumor (ET) on the BraTS 2020 dataset of 0.8357, 0.8660, and 0.8004, respectively.
average dice score of 0.883 and a Hausdorff distance of 7.84 on Brats 2021 cross validation.
method
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
ResNet-101, and DenseNet-121 for tumor segmentation in detail and used 3D slicer tools for MRI imaging. Eyad Gad [4] explored an attention-based 3D UNet model for segmentation and claimed to have improved the segmentation by concentrating on useful portions of the image. IKE Purnama [5] used a modified U-Net and claimed to have improved detail and accuracy in the segmented images. Zhao J [6] proposed Uncertainty-aware Multi-dimensional Mutual Learning. Gayathri Ramasamy [7] constructed a multi-modal modified Link-Net model utilizing Squeeze and Excitation ResNet152 as a segmentation backbone for improved segmentation performance. Cihan M [8] applied convolutional neural networks for fusion of phase MRI images: pre-contrast, venous, arterial, and delayed phases. With all these advancements, there are still limitations on the applications of research results. Some studies, for instance, report an OS prediction accuracy of merely 44.8%, illustrating ongoing difficulties in achieving dependable clinical results. A. Durgut [9] studies image processing using parallel and serial programming methods and finds parallel programming reduces the time required for processing both whole and segmented images. Whole image processing outperformed segmented image processing. The research highlights the efficiency of parallel programming in improving image processing tasks. Table 1 provides further insights into related research contributions and their respective metrics. Using nature-inspired optimization methods such as genetic algorithms as well as hybrid filters can provide solutions to the issues of brain tumor image analysis[18]. The improvement in the use of these approaches provide methods for optimizing hyperparameter parameters of a model, reduce the effects of over fitting and improve the original feature extraction to enhance model accuracy and generalizability. Through mimicking natural behavior and incorporating heterogeneous filtering methods, these approaches present a valuable strategy to enhance productivity and applicability of machine learning models for use in clinical settings.
Materials And Methods
The issue we address in this work is that of accurate prediction of overall survival (OS) in glioblastoma multiform (GBM) patients that are undergoing stereotactic radio surgery (SRS). GBM is known for its aggressive nature and poor prognosis which in turn sees the rise of SRS as a key treatment modality for precise tumor cell targeting. That said accurately prognosticating OS in GBM patients post SRS is still a challenge which we put down to the heterogeneous nature of the disease and also to the variable patient responses to treatment. Also included are the goals of our work. • Preprocessing the Dataset for feature extraction and classification by applying a Hybrid Bilateral_Gaussian
Filter, and normalization to reduce noise, and improve performance. Segmentation for tumor by Cat Swarm Optimization for Medical Image Segmentation (CSO-MIS) assisted hybrid ResNet and U-Net deep learning model. Ensemble of popular machine learning algorithms— Random Forest, KNN models, and SVM—as well as the deep learning model ANN are employed to predict the OS. To evaluate the performance. Compare accuracy, sensitivity, specificity, and area under the ROC curve metrics
Implementation Figure 1 showcases the remarkable process of Glioma sub region segmentation and predicting overall survival in glioblastoma patients using deep learning architecture. First out of the gate we do in depth pre processing of the BRATS 2020 dataset which we do to that very meticulous degree which we do for the sake of accuracy. Then we do segmenting which in turn does what it can to improve the accuracy of the results. Also we put the data into train, test, and validate sets which we do to make our analysis as through as possible. We see this approach to be very effective in that it puts out very high in terms of accuracy and reliability in what we predict for overall survival in patients with glioblastoma. Dataset As part of an annual challenge related to the segmentation of brain tumors using multimodal MRI scans, a set of datasets called brain tumor segmentation (BraTS) 2020 has been made available. The BraTS 2020 dataset contains glioma patient magnetic resonance imaging (MRI) scans from different medical facilities. The participating institutions are focused on glioma scans. To ensure accuracy and reproducibility of results, standardization of protocols has been done. The dataset also contains expert annotated delineations for various tumor subregions, which are essential for training supervised learning models. The delineations are demonstrated through: • Label 1: Non-Enhancing Tumor Core (NCR/NET) • Label 2: Peritumoral Edema (ED) • Label 3: GD-enhancing Tumor (ET) For each case, clinical information such as the age of the patient is provided. Survival data, i.e., the number of days the patient survived post-initial diagnosis, might also be available for a subset of patients. Data Preprocessing To preprocess BraTS2020 MRI dataset, performed a sequence of normalization followed by application of hybrid filter for noise reduction. This approach ensures that the data is scaled properly and noise is reduced, which are crucial steps for improving the performance of downstream tasks like segmentation and classification. A crucial first step in data analysis, particularly in the domains of machine learning and image processing, is normalization.
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
This includes setting the data values onto a common scale so we can stabilize training processes and improve algorithm performance. Normalization can account for variability in intensity from medical images (e.g., MRI scans) that might occur due to the use of different scanning protocols, or machine calibration during scans from other institutions. This process generally rescales image intensity values of MRI scans to a standard range of [0, 1] or [-1, 1], which enhances image comparison and analysis. Perform
the normalization and then apply a Hybrid filter as given in Algorithm 1 to each slice to reduce noise. As given, the Hybrid filter reduces noise within the MRI images while improving the overall image quality with local analysis of pixel neighborhoods. This normalization and noise reduction with the Hybrid filter prepares the BraTS2020 MRI dataset for more effective analysis in segmentation models or other machine learning tasks by fostering higher quality images and greater consistency throughout the dataset.
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
HybridBilateral_Gaussian Filter Algorithm1: Hybrid Filtering (Bilateral_Gaussian Filter) 1 2 3 4 5 6
Input: Original image org(𝑖,𝑗) Gaussian sigma value 𝜎 for Gaussian filtering
Bilateral Filtering Apply bilateral filtering to the org image to result a Denoised image Denoised img: org_denoised(𝑖,𝑗)=denoise bilateral (org(𝑖,𝑗))
Gaussian Smoothing Apply Gaussian smoothing to the org image with the specified sigma value to obtain a smoothed version. Smoothed image: org_smoothed(𝑖,𝑗)=gaussian_filter(org(𝑖,𝑗),𝜎)
Hybrid Filtering For each pixel (𝑖,𝑗)(i,j) in the image: Calculate the hybrid filtered pixel value as the sum of the Denoised pixel value and the difference between the org pixel value and the smoothed pixel value.
Hybrid filtered pixel value: hybrid_filtered(𝑖,𝑗)=org_denoised(𝑖,𝑗)+(org(𝑖,𝑗)−org_smoothed(𝑖,𝑗)) Output: Return Hybrid filtered image: hybrid_filtered(𝑖,𝑗)
Model The suggested work uses a ResNet encoder and a U-Net decoder which captures the advantages of both networks. This fused method uses the efficient feature extraction of ResNet and the segmentation characteristics of U-Net to achieve feature extraction and spatial information preservation concurrently. The deep layers of ResNet improve the fine details retrieval, and U-Net’s encoder-decoder configuration accurately captures the details in precise segments. The residual connections of ResNet also address overfitting to improv generalization. This enhances the model’s strength on the variability and complexity of the medical images, which make it difficult to analyze. This fusion has demonstrated state-of-the-art performance across segmentation tasks, offering a powerful and robust solution for accurate tumor segmentation. Resnet as Encoder U-Net As Decoder To sufficiently segment brain tumors within the BraTS2020 dataset using the strengths of both ResNet and U-Net architectures, it is important to comprehend how ResNet may serve as an encoder in a U-Net framework and how this bestows certain advantages, along with its influence on the overall functionality of the network. The ResNet-Unet Architecture is represented in Figure 2.
The encoder portion of U-Net, harnesses ResNet’s capacities for deep learning while the U-Net architecture allows for ample amount of local context with its skip connections. This allows for ResNets reliable feature extraction via residual blocks, particularly with deeper architectures such as ResNet-50 or ResNet-101. Each residual block is fundamentally composed of two parts: several convolutional layers that are viewable as executing the main operation; and a skip connection, which sums the block input with the block output. This particular arrangement solves the issue of vanishing gradients, expediting the ability to train deeper network architectures without degradation in performance because of sufficient gradient flow. U-Net in both its fundamental approach and framework for model development as well. The methods given above are highly specialized for brain tumor segmentation and their development needed expertise and extensive experimentation [19]. The network gains from the ability to encode more abstract and complicated information from the input images when ResNet is used as the encoder in U-Net. This is especially useful for medical imaging jobs where it can be critical to discern minute distinctions between tumor types and normal tissues, such as tumor segmentation in MRI scans. ResNet’s improved depth and feature extraction capabilities enable the network to collect
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
the finer features and the larger context required for precise segmentation. Mathematically, a residual block in ResNet can be represented as in equation 1: Output = ReLU (BatchNorm (Conv (x))) + x
Where: x is the input to the block, Conv denotes a convolutional layer, Batch Norm represents batch normalization. ReLU is the rectified linear unit activation function and +x represents the skip connection. Several residual blocks stacked together make up the encoder part of the network. The encoded representation is the output of the last residual block. In the U-Net decoder, the segmentation map is generated from the encoded features through the sequential processes of up convolution and concatenation. Up convolution (also termed transpose convolution) is a process in the decoder that enlarges the spatial dimensions of the feature maps. These up sample maps are then concatenated with the same connected feature maps from the encoder, via the skip connection.This procedure preserves spatial hierarchies, which aids in correct localization, and enhances the feature maps with fine-grained information required for exact segmentation.Mathematically, the up sampling operation can be represented as in equation 2 : Upsampled = Conv(Upsample(x))
Where x is the input to be up sampled, up sample denotes the up sampling operation, Conv represents a
convolutional layer. Then, to incorporate skip connections, we concatenated the corresponding feature maps from the encoder with the up sampled feature maps in the decoder. The combined feature maps are then processed through additional convolutional layers to refine the segmentation mask. Optimization In this study we present our put forth of Novel Cat Swarm Optimization for Medical Image Segmentation (CSO-MIS). This Cat Swarm Optimization which we have used improves medical image segmentation by fine tuning the parameters of the deep learning models we use for tumor segmentation. We apply CSO to optimize the hybrid ResNet and U Net architectures which in turn improves their performance in accurate tumor detection. By tuning hyperparameters and network settings we see an improvement in segmentation accuracy and efficiency which in turn leads to better identification of key tumor areas. In this algorithm cats we see exploration of promising solutions and at the same time we have elements of random exploration which together do a good job at the balance of exploration and exploitation to find high quality segmentations. Our algorithm has cats which over time play in and out of the solution space to do what it takes to find the best segmentations of the medical image..CSO is inspired by the hunting behavior of cats and their ability to cooperate to solve complex tasks. Here’s a simplified version of how CSO can be applies for this implementation:
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
Algorithm 2: Cat Swarm Optimization for Medical Image Segmentation (CSO-MIS) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 19 20
Input: Medical image img, Number of cats 𝑁m, Maximum iterations 𝑇, Learning rate 𝜂, Exploration rate 𝜖 Output: Segmented image 𝑆eg∗ Procedure: Initialization: Randomly initialize 𝑁m cats in the solution space representing potential segmentations of the image. Set best solution found so far 𝑆eg∗=None and best fitness best_fitness=∞. Main Loop: for 𝑡=1 to 𝑇 do: Update each cat: Move each cat towards the best solution found so far with a probability of 𝜂. else, randomly explore the solution space with a probability of 𝜖. Evaluate the fitness of each solution (segmentation). Update best solution: if any solution has a better fitness than best_fitness then: Update best_fitness and 𝑆eg∗ with the best-performing solution. Segmentation Inference: Use the best solution 𝑆eg∗ obtained from CSO to segment new medical images. Output: Segmented image 𝑆eg∗. End
Randomly initialize a cats population representing different sets of model parameters. for each iteration, Evaluate the fitness of each model (cat) using the fitness function and the training dataset. The fitness function is as given in equation 3. (3) Where: S is the predicted segmentation mask, G is the ground truth segmentation mask, |S∩G| is the number of overlapping pixels between S and G, |S| and |S| are the total number of pixels in S and G respectively. Update the present best solution based on fitness values. Update the position of each cat using the rules of CSO which include both exploitation (movement toward the best solution) and exploration (random search). The new position xi(t+1) of the i-th cat at time t+1 is given by the equation. 4: xi(t+1) = xi(t)+η⋅(global_best−xi(t)) + 𝜖⋅rand()⋅(upr_bound−lwr_bound)
Where: At time t the position of the i-th cat is x_i , and 𝜂 is the learning rate. Global_best is the best solution to date, 𝜖 is the exploration rate. rand () returns a random number between 0 and 1. Upr bound and lwr bound define the solution space. Update the best solution found so far based on the fitness of the cats’ positions. If any Cat finds which set of
parameters performs better for improved segmentation results, report that as the best solution at present. Do this for a fixed number of iterations or until we see that the model has converged. Once the optimization process is done, train the segmentation model with the best set of parameters found by the CSO on the full training dataset. Then put the trained model to the test on a separate validation set to see how it does. This step also makes sure the model does well on data it has not seen before.
Results And Discussion
The in Figure 3 we present a ROC curve which reports on a classification model’s performance across three different classes. We plot True Positive Rate (sensitivity) against False Positive Rate which in turn gives us info into how the model does at different threshold levels. A curve that reaches towards the top left hand corner of the plot indicates better performance which in turn means high sensitivity and low false positives. By looking at the three curves we are able to see which class the model does best at identifying as well as the models’ overall performance which in turn may point out areas for improvement. Also if we see large deviations from the ideal curve that may indicate issues with the models’ predictions in certain areas which in turn may require more refinement. The dice coefficient for both the training and validation datasets over 50 epochs is presented in Figure 4. The plot represents an overall upward trend, indicating that segmentation accuracy improves as training progresses. This suggests that the model is learning effectively and becoming more accurate with each epoch for both the training and validation data.
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
Figure 5 shows the training and validation accuracy (left) and loss (right) over 50 epochs. From the accuracy curves showing the training and validation group data, we see that both are shifting in a consistent upward direction, suggesting that the model is improving in making accurate predictions over time. We also see that the loss curves are decreasing in a consistent manner, indicating that, as training continues, the model’s predictions are making fewer mistakes (errors) with each epoch. The general stability in the training and validation accuracy and loss curves suggests that the model is learning to generalize from the training data to the validation data with no signs of overfitting, which improves overall model performance for segmentation tasks. Figure 6 shows the Training and Validation Mean Intersection over Union (IOU) over 50 epochs, where
the X-axis represents the epochs and the Y-axis shows the Mean IOU score. Both the Training and Validation Mean IOU scores show a steady, nearly identical increase, indicating consistent performance and minimal overfitting. The values approach high accuracy in image segmentation. The final stage of the U-Net structure assigns each pixel a distinct class label ( an example is shown in Figure 7 with labels including “NOT tumor,” “NECROTIC/CORE,” “EDEMA,” and “ENHANCING”). While it is the U-Net decoder that is providing precise localization of the features, ResNet encoder is providing distinct feature extraction providing a segmentation (i.e., class) map that preserves high resolution similar to the training dataset and test ground truth. The segmented brain tumors additionally classified into separate classes: i.e., core, edema, and enhancing
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
tumor were represented collectively as the ground truth, and represent a comprehensive view of the tumor classes. This way of representing tumor classes will ultimately assist in data accuracy and reliability concerning deep learning based brain tumor segmentation. The 99.2% segmentation accuracy in this study shows a significant improvement over common Glioblastoma segmentation methods due to the tumor’s complexity. This improvement is a function of the hybrid ResNet-UNet that leverages ResNet’s feature extraction and U-Net’s good spatial accuracy, which leads to the segmenting of the glioma accurately. The Hybrid Bilateral-Gaussian Filter is also instrumental in helping reduce noise and at the same time maintain the edges that are important for the model not to receive garbage inputs into the network. Also, CSO helps the model with hyperparameter tuning, which optimized ResNet and U-Net for better performance. Overall, these techniques (ResNet-U-Net, CSO, Hybrid BilateralGaussian Filter) are better than existing approaches which means greater precision in Glioblastoma segmentation and better clinical decision-making. Survival Prediction The BraTS2020 challenge requires the use of a dataset that includes features obtained from segmented MRI brain images, patient age, and surviving days for survival analysis [20]. The aim is to predict overall survival (OS) throughout intervals like short--, mid-, and long-term. To predict the Overall Survival time, radiomic and image-based features were extracted. These features have been shown to be successful in different studies for OS tasks [21]. In order to approach the task, we used a process of feature extraction and engineering combined clinical data and the derived imaging features that describe the tumor size and region [22]. After which we trained a number of machine learning models e.g., frequently used methods-support vector
machine (SVM), random forest, k-nearest neighbors, and artificial neural networks in the combined dataset. [23]. The first step in the procedure is gathering data, which includes patient demographics such as age, specific survival durations, and MRI images that are divided into different tumor locations (necrotic/core, edema, enhancing). We combined the volumetric and location features with the patient’s age to train a model using linear regression [24]. Quantitative properties of the tumor areas, including their size, shape, and intensity, are extracted from these MRI scans and are thought to have an impact on survival rates. Features can generate prognostic imaging signatures for OS prediction and patient stratification for GBM [25]. Which are very effective with a small number of training data? Random Forest is an ensemble algorithm that is made up of multiple decision trees and operates on them as an ensemble. It is one of the most frequent [26]. Machine learning models like Random Forests, Support Vector Machines, or Neural Networks are then trained on this integrated dataset, which combines clinical data with imaging-derived features. To estimate how long patients with similar profiles will live after diagnosis, the models are trained to correlate the features with survival durations. All models were fit with knowledge that the models were finally tuned and evaluated with methods such as cross-validation to provide a safeguard against overfitting and optimize hyperparameters. All models were measured with accuracy, precision, recall, F1-score, and area AUC-ROC. These statistical methods provided a capacity for predicting survival with a compelling analysis that used advanced analytic techniques to improve health outcomes on a patient level. To visualize, Figure 8 provides a pictorial of survival rates. The x-axis is days survived, and the y-axis is people. This relationship with features may allow some associations to be developed in survival rates that may assist in developing
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
beneficial treatment for variations in patients with various medical conditions. Figure 9 provides insight into the dependency of survival data and segmentation, determining whether the survival rate is more dependent on the core region, edema, or enhancing tumor. Based on the image, it appears that the survival rate depends largely on the edema region. This information can be useful in guiding the development of targeted treatment strategies for patients with glioblastoma. As we evaluated predictive survival models, we explored a range of supervised machine learning models, which included random forest classifier, SVM, KNN, and ANN. Random forest produced the highest testing accuracy (0.54) and precision (0.8) but had a relatively low training score (0.37), which can be an indicator of a model that is not learning from the training data as we want. The other models we explored (i.e. the SVM, KNN, and ANN models) had similarly low performance, especially SVM and KNN, and overall had lower accuracy and precision than the random forest model for testing. We used the three best models, Random Forest (RF), Artificial Neural Network (ANN),
and support vector machine (SVM) to then enhance the overall prediction accuracy of survival using an ensemble strategy see Figure 10. The ensemble model is here to be defined: • Random Forest (RF): Random Forest is an ensemble model that builds multiple decision trees and combines the model outputs, so it can model complex interactions between variables that single models may be unable to decode. Random Forest also provides feature importance scores that allow for the evaluation of which factors contribute most to the survival outcomes, which is important for clinical interpretation and personalized treatment models. • Artificial Neural Network (ANN): ANN builds on layers of nodes so that identifier complex, non-linear patterns are possible in large datasets. • Support Vector Machine (SVM): SVM finds the optimal hyperplane for dividing classes while performing especially well in higher dimensional spaces. The Ensemble method (or approach) takes advantage of the unique strengths of each model (where each model has
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
unique strengths: Random Forest makes robust, low-variance predictions; ANN can model highly complex non-linear relationships; SVM can handle high-dimensional spaces). The power of the Random Forest algorithm is its ability to capture the complex relationships between the array of different features, which is particularly useful when working with high-dimensional medical data (for example, the imaging-derived features from GBM). However, the Random Forest and other models could not combine and capture all of the patterns in the data, which is what the ensemble method was intended for. Each model captures its own features of the available data and reduces the overall errors in prediction because the weaknesses of each individual model are reduced by other models. However, by ensemble methods, predictions are based on a combined model with the strength of each individual model, providing more stability and robustness by dampening individual model biases and variability. The ensemble method provides improved prediction accuracy and reliability and could be particularly advantageous for the difficult survival
prediction for the cohort of Glioblastoma patients. The overall predicted value for an input given a model can be represented as follows Ensemble Prediction= (WRF .PRF)+(WANN.PANN) +(WSVM.PSVM)
Where: WRF, WANN and WSVMare the weights assigned to the RF, ANN, and SVM models, and PRF, PANN and PSVM are prediction probabilities of RF, ANN, and SVM models respectively. The confusion matrix for the survival prediction is represented in Figure 11. Table 2 displays several different models for prediction of survival outcomes with metrics, including accuracy, precision, training score and testing score. The Ensemble model, which attained the highest accuracy of 60.01% out of all models presented, is the best model for this type of data. Moreover, the Ensemble model achieved a precision score of 0.85 meaning that it is better at correctly predicting positive survival outcomes when present. Its model training score of 0.63 and model
SVM: Support vector machine; KNN: K-nearest neighbors; ANN: Artificial neural networks.
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
testing score of 0.60 indicate that the model generalizes sufficiently to new or unseen data making it an ideal model for survival prediction tasks. The Figure 12 compares the accuracies of segmentation and survival predictions made by different models. Each bar represents a different model. The survival prediction accuracy figure indicates that our models have different levels of correspondence with survival outcomes, with attainment of the highest accuracy from predicted survival, while other models performed on a moderate level indicating that some potential exists for improving survival predictions. In the segmentation prediction accuracy figure, our models were shown to be superior, attaining the highest accuracy, while the other models had lower accuracy. Thus, the other models may be less able to perform well with precise segmentation tasks. This study differs from typical clinical management methods of survival prediction in Glioblastoma patients using BraTS2020 dataset in that, all we did was create an ensemble of advanced machine learning models (i.e., Random Forest, ANN, and SVM) to analyze all of the complex imaging data in a more sophisticated manner than just clinical features and subjective manual interpretation. Typical clinical management of glioblastoma patients often relies on clinical features from prior research and clinical decision support systems to inform their suggested management. This approach may be subjective and the reliability is somewhat less refined. The study’s approach used an advanced ensemble of models which captured complex data patterns found within the dataset and produced a more objective and precise prediction system. The added benefits of this approach include the ensemble produced a much better accuracy (60.01%), it generalized better to new data, and has the potential to enhance clinical decision making by having a more reliable prediction system for survival during clinical management of Glioblastoma patients.
Conclusion
Glioblastoma is the most common and aggressive form of primary brain tumor in adults that poses significant clinical management difficulties including high rates of recurrence, poor prognosis, and complexity and heterogeneity. This study aims to improve predictions of overall survival for Glioblastoma patients treated with stereotactic radiosurgery using advanced image segmentation and machine learning. We used Cat Swarm Optimization assisted hybrid ResNet and U-Net architectures to perform precise tumor segmentation and additional machine learning algorithms as an ensemble to make strong predictions of survival time. Utilizing the BraTS2020 dataset, we accurately delineated three critical tumor regions: core, edema, and enhancing tumor. The segmentation model achieved remarkable metrics, including a segmentation accuracy of 99.2%, a loss of 0.023, a recall of 0.986, a mean intersection over union (IOU) of 0.991, a dice coefficient of 0.96, a precision of 0.991, a sensitivity of 0.991, and a specificity of 0.997. In this study, we evaluated a number of machine learning models for survival prediction, and prioritized using different machine learning models for survival prediction while the Random Forest algorithm was ranked the best tool for survival prediction post deriving the features from segmented images due to its ability to handle the complexity of the features. Although the ensembling approach for survival prediction at 60.01% improved accuracy, these results demonstrated that there is a lot of potential to improve survival prediction for Glioblastoma patients in terms of clinical management and prognosis by utilizing advanced image segmentation methods with machine learning models. In future work using the BraTS dataset, advanced machine learning models, particularly Transformers to process spatial data should be employed. In addition, we can extend the use of multimodal data, which may include genomic profiles and comprehensive medical history, to improve prediction accuracy and provide personalized treatment strategies. Improving
Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
methods for data handling that can ensure privacy could result in a more broaden data demographic for a training dataset dataset. Performance and randomized controll trails in real-time clinical decision support systems could foster patient outcomes helpful toward personal medicine management.
Acknowledgments
We extend our sincere thanks to all researchers and contributors in the field of glioma research. Their pioneering work has greatly inspired and guided this study.
Data Availability Statement
The authors confirm that the data that supports the findings of this study are available within the article. Raw data that support the finding of this study are available from the corresponding author, upon reasonable request.
Conflict Of Interest
The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Ethics
There are no ethical issues with the publication of this manuscript.
Statement On The Use Of Artificial Intelligence
Artificial intelligence was not used in the preparation of the article.
References
- scans using U-Net. SN Comput Sci 2023;4:386.
- Rajput S, Kapdi RA, Raval MS, Roy M, Bhalodiya [CrossRef] JM. Deriving and interpreting robust features for [14] Sangui S, Iqbal T, Chandra PC, Ghosh SK, Ghosh A. survival prediction of brain tumor patients. Int J 3D MRI segmentation using U-Net architecture for Imaging Syst Technol 2024;34:e23105. [CrossRef] the detection of brain tumor. Procedia Comput Sci
- Hong S, Wu J, Zhu L, Chen W. Brain tumor classifi- 2023;218:542–553. [CrossRef] cation in VIT-B/16 based on relative position encod- [15] Soltaninejad M, Pridmore T, Pound M. Efficient ing and residual MLP. PLoS One 2024;19:e0298102. MRI brain tumor segmentation using multi-reso- [CrossRef] lution encoder-decoder networks. In: Med Image
- Phan TC, Phan AC, Nguyen KT, Tran HD. Detection Comput Comput Assist Interv. Cham: Springer; and segmentation of brain tumors on 3D MR images 2020. p.31–39. [CrossRef] using 3D U-Net. In: Future Data and Security [16] Isensee F, Jäger PF, Full PM, Vollmuth P, Maier-Hein Engineering. Singapore: Springer Nature; 2023. p. KH. nnU-Net for brain tumor segmentation. arxiv. 528–541. [CrossRef] org/abs/2011.00848 doi: 10.48550/arXiv.2011.00848 1606 Sigma J Eng Nat Sci, Vol. 43, No. 5, pp. 1592−1606, October, 2025
- Ali MJ, Akram MT, Saleem H, Raza B, Shahid AR. [22] Bhagyalaxmi K, Dwarakanath B, Reddy PVP. Deep Glioma segmentation using ensemble of 2D/3D learning for multi-grade brain tumor detection and U-Nets and survival prediction using multiple fea- classification: a prospective survey. Multimed Tools tures fusion. In: Med Image Comput Comput Assist Appl 2024;83:65889-65911. [CrossRef] Interv. Cham: Springer; 2020. p.179–188. [CrossRef] [23] Saxena S, Agrawal A, Dash P, Jena B, Khanna NN,
- Das S, Bose S, Nayak GK, Satapathy SC, Saxena S. Paul S, et al. Prediction of O-6-methylguanine-DNA Brain tumor segmentation and overall survival pre- methyltransferase and overall survival of glioblas- diction in glioblastoma multiforme using radiomic toma patients using MRI-based hybrid radiomics features. Concurr Comput Pract Exp 2020;32:e6501. signatures in machine and deep learning frame- [CrossRef] work. Neural Comput Appl 2023;35:13647–13663.
- Wang S, Dai C, Mo Y, Angelini E, Guo Y, Bai W. [CrossRef] Automatic brain tumour segmentation and biophys- [24] Babaei R, Bonakdar A, Shakourifar N, Soltani M, ics-guided survival prediction. In: Crimi A, Bakas Raahemifar K. Evaluating maximum diameters of S, editors. Brainlesion: Glioma, Multiple Sclerosis, tumor sub-regions for survival prediction in glio- Stroke and Traumatic Brain Injuries. Lect Notes blastoma patients via machine learning considering Comput Sci 11993. Cham: Springer; 2020. p. 29–41. resection status. J Electron Inf Syst 2024;6:22–38. [CrossRef] [CrossRef]
- Das S, Swain MK, Nayak GK, Saxena S, Satpathy Preoperative computed tomography-based tumoral SC. Effect of learning parameters on the perfor- radiomic features prediction for overall survival in mance of U-Net model in segmentation of brain resectable non-small cell lung cancer. Front Oncol tumor. Multimed Tools Appl 2022;81:34717–34735. 2023;13:1131816. [CrossRef] [CrossRef]
- Hajianfar G, Haddadi Avval A, Hosseini SA, Nazari Agarwal A, Priya S. Clinical and magnetic resonance M, Oveisi M, Shiri I, et al. Time-to-event overall imaging radiomics-based survival prediction in glio- survival prediction in glioblastoma multiforme blastoma using multiparametric magnetic resonance patients using magnetic resonance imaging radiom- imaging. J Comput Assist Tomogr 2023;47:919–923. ics. Radiol Med 2023;128:1521–1534. [CrossRef] [CrossRef]
Share and Cite
K, B.; B, D. Cat swarm optimized tumor segmentation and an ensemble predictive model for glioblastoma patient sur. Sigma Journal of Engineering and Natural Sciences 2025, Vol. 43, pp. 1592-1606. https://doi.org/10.14744/sigma.2025.00029