The LGG Segmentation Dataset is a publicly accessible database of MRI brain scans for lower-grade gliomas, originally released by Mazurowski et al. (6). It involves MR images with manually provisioned FLAIR abnormality segmentation masks that allow for the exact delineation of chromatic abnormality regions. The dataset includes data from 110 patients from The Cancer Genome Atlas (TCGA) LGG collection, all of whom have FLAIR sequences and associated genomic cluster information. In addition to the images, patient metadata and genomic tumor cluster information are included in a separate CSV file for integrated studies of imaging and genomic characteristics. The MRI scans are from The Cancer Imaging Archive (TCIA), providing a set of standardized and publicly available imaging data. This data has been used to investigate the associations between tumor shape characteristics, as determined by deep learning, and a tumor’s genomic subtype, as well as the effect size of these associations on patient outcomes. Figure 2 shows an example from the LGG Segmentation Dataset.

Preprocessing medical images is of utmost importance to maintain data in a standardized form, ensuring all properties of robustness and suitability for deep-learning models. In PoSAM-ULTRA, the entire pipeline of preprocessing steps was designed, encompassing resizing, normalization, data augmentation, and dataset partitioning, of MRI scans and their respective tumor masks of the LGG dataset (7). These steps ensure standard input dimensions, minimize intensity variability between subjects, increase training examples, and allow for unbiased evaluation. Resizing since an MRI scan naturally differs in resolution and aspect ratio, both images and their respective masks were resized to a uniform resolution of 256 × 256 pixels guaranteeing compatibility of all inputs with the encoder-decoder-style PoSAM-ULTRA, while simultaneously providing adequate spatial detailing for precise detections of tumor boundaries. I represent the original MRI image (8). One gets the corresponding ground mask resized with the same transformation to maintain spatial alignment, as shown in Equation 1.

Normalized medical images often suffer from varying intensity distributions due to differences in acquisition protocols and patient-specific contrast variations (9). To stabilize the training process and accelerate convergence, pixel values were normalized to the range [0, 1] using min-max scaling I_min and I_max denote the minimum and maximum pixel intensities of the image, respectively. This normalization ensures that the input dynamic range is consistent across all subjects, allowing the network to focus on structural differences rather than intensity biases, as shown in Equation 2.

Data augmentation to improve generalization and reduce overfitting, extensive on-the-fly augmentation was applied during training (10). Augmentation increases the diversity of the dataset by simulating variations in patient anatomy and acquisition conditions. The following transformations were randomly applied. Rotation random angular rotations up to ± 20, simulating changes in patient orientation. Flipping: Random horizontal and vertical flips; please enforce symmetrical invariance (11). Scaling/Zooming small random rescaling of images, making the training robust toward changing tumor sizes. Intensity Adjustment changes brightness and contrast to simulate changes in imaging conditions. Formally, the augmented dataset is defined as D is the original dataset, M is the mask, and 𝒯 is the distribution of transformations applied. This guarantees that the model sees more diverse tumor appearances, which is critical for robust segmentation of low-grade gliomas, as shown in Equation 3.

The dataset was divided into 80% training (3,143 images), 10% validation (393 images), and 10% testing (393 images) to ensure a fair and unbiased assessment of performance. The training set D_train is used to optimize model weights via backpropagation. The validation set D_val is used for hyperparameter tuning and observing model convergence. The test set D_test is reserved for the final evaluation to get a measure of conscious ability as shown in Equation 4.

To generate the four-channel inputs for the SAM backbone, a prior-information map is incorporated to highlight glioma probability regions. This prior-map is generated via a lightweight preprocessing pipeline that begins with skull-stripping using HD-BET and image intensity normalization using a Z-score. A primary tumor probability map is then produced through unsupervised clustering (K-means with $k = 3$), where the cluster with the highest average intensity is considered the primary tumor candidate. This mask is refined using morphological closures and a median filter to reduce noise, and then min–max normalization to fit the range [0, 1]. This prior-channel approach has proven to provide meaningful spatial semantics, confirming that it provides additional structural cues that identify potential tumor regions without replacing the detailed learning process of the model.

A brain tumor dataset is, generally, characterized by extreme class imbalance, for tumor-region image pixels make up a much smaller proportion of the total image than do background pixels. To confront this, the weights of the loss function were optimized through PBFO to maintain an equilibrium between Dice loss and binary cross-entropy (BCE). Thus, the model is made to be sensitive to the tumor of small structures, while at the same time avoiding any chance of its overbearing influence by background pixels. The preprocessing pipeline allowed for the standardization, diversification, and balancing of the dataset from which PoSAM-ULTRA can draw worthy inputs. These would have helped directly in accurate tumor boundary detection, patient-to-patient generalization, and reduction of overfitting of the model so that it can perform well in segmentation against the baseline methods.

The method integrates Segment Anything Model (SAM) as a backbone with a ResNet34-based 4-channel encoder, enhanced by CBAM attention modules, extended skip connections, and deep supervision through auxiliary outputs. To optimize performance, we employ the Polar-Bear Foraging Optimization (PBFO) algorithm for efficient hyperparameter search (Figure 3).

PoSAM-ULTRA applies the Polar-Bear Foraging Optimization (PBFO) algorithm to fine-tune hyperparameters vital to the system such as the learning rate, batch size, dropout ratio, weight decay, and the number of filters per encoder block (26). Each candidate hyperparameter configuration is encoded as a vector, where h_ij stands for the j-th hyperparameter (Figure 4).

Given iteration t, the population of candidate solutions is denoted; altogether, there are m candidates. This initial population is randomly generated within a predefined bound. U(a,b) stands for the uniform distribution (27). For instance, the learning rate may be picked in the range [10⁻⁵,10⁻²], batch size in the range [8,64] and dropout rate in the range [0.1,0.6]. Each candidate solution then undergoes evaluation against a fitness function that reflects the quality of the segmentation. For PoSAM-ULTRA, the Dice similarity coefficient is primarily used as a metric, as shown in Equations 23–26.

Y^⁢(Hit) represents the forecasted segmentation resulting from training the model with hyperparameters Hit. This enhances segmentation accuracy during optimization by ensuring that candidates that have high segmentation accuracy will be selected (26). PBFO alternates between a global roaming phase of exploration and a local hunting phase of exploitation. The exploration stage gives polar bears the chance to search through large portions of hyperparameter space to minimize the risk of premature convergence. The update of candidate solutions is defined as follows, as shown in Equation 27.

Here, Hbextt refers to the best solution discovered thus far, and Hjt and Hkt refer to two randomly chosen distinct solutions. Vectors of scale, R₁,R₂U(0,1), are sampled from a Uniform (0,1) distribution, with parameters α,β controlling the degree of exploration. The definition emphasizes the counterbalance between convergence to the best solution and the diversity created due to interaction between individuals in the population. When an interesting area is located, PBFO proceeds toward exploitation, imitating local hunting and improving candidate solutions about present best hyperparameters (28), controlled by the parameter γ ∈ (0,1), which denotes a fine degree of adjustment on local search. These three parameters, α,β, and γ, dynamically balance exploration and exploitation over iterations; exploration is prominent during the initial iterations (with α,β large), and, as the later iterations proceed, exploitation begins to dominate (with γ increasing), thus preventing premature convergence and enhancing the chances to find globally optimal hyperparameter. The optimization terminates when the number of iterations reaches T_max or when improvement in fitness becomes lower than some threshold ∈, and the last optimal hyperparameter configuration is given by, as shown in Equations 28–30.

With the help of PBFO demonstrated in Table 1, PoSAM-ULTRA meaningfully searches the hyperparameter space, speeds up model training, decreases computation cost, and improves the segmentation results on lower-grade glioma MRI datasets. This principled strategy assures the model will achieve robust learning and a high level of accuracy without any manual trial-and-error.

To put PoSAM-ULTRA to a thorough test, we have run comparative experiments against three widely used brain tumor segmentation models: U-Net, U-Net++, and nnU-Net-lite. It is worth mentioning that these are all baseline models, and they have had their hyperparameters tuned by the Polar-Bear Foraging Optimization (PBFO) algorithm. Hence, while they served as competing baseline methods, all their performances were evaluated under the best parameter settings, making the final comparison fair and unbiased.

U-Net is a classical encoder-decoder network with symmetric skip connections (29). While it allows efficient segmentation, it simply lacks any attention mechanism and multi-modal feature integration, rendering it incapable of capturing refined tumor boundaries.

U-Net++ suggests that nested skip connections and dense skip connections improve gradient flow and multi-scale feature fusion (30). However, since the architecture “squeeze” is reliant on nested skip connections, it is possible that this architecture does not emphasize areas related to tumor-specific regions in lower-quality MRIs.

nnU-Net-lite is an adaptive lighter version of nnU-Net that automatically adjusts the layer depth of the architecture to meet the dataset’s requirements (31). With PBFO based hyperparameter optimization, the nnU-Net achieves a solid baseline performance, but the model’s architecture will prefer low computation given the shallowness of the network and the attention implemented within the model may not fully resolve lower contrast sections of the MRIs that require high computation use.

PoSAM-ULTRA and the baseline networks (U-Net, U-Net++, and nnU-Net-lite) were trained on LGG, with hyperparameters optimized via the dynamic Polar-Bear Foraging Optimization (PBFO) technique. PBFO instantiates the best learning rate, batch size, optimizer type, and loss weighting to allow for fair and correct comparisons between models. The dataset was split into 80% training, 10% validation, and 10% test datasets. Preprocessing included resizing and normalization in addition to data augmentation (rotation, flipping, scaling, intensity variation) to increase robustness and minimize overfitting. At training time, PBFO manages the optimization of the weighting between Dice loss and binary cross-entropy to handle the class imbalance of tumor and non-tumor regions. It propagates early stopping and checkpoint saving for the best validation performance. For each evaluation metric, values were computed on the test set to measure segmentation and tumor localization performance. Dice Coefficient measures the overlap between the regions of tumors predicted and the ground truth. Very high values of Dice represent very accurate segmentations of tumor boundaries, which is important for accurate localization (32). Intersection over Union quantifies the measure of intersection and union of predicted and true tumor regions (33). This complements Dice by giving a negative score if the prediction over-segments the tumor; thus, it is useful to assess if the segmentation covers the tumor very precisely, as shown in Equations 31, 32.

Accuracy measures the percentage of total pixels correctly classified throughout the entire image and therefore provides insight into a model’s intrinsic accuracy in distinguishing tumors from non-tumor tissues. Precision is the percentage of pixels predicted as tumors that are tumors. When precision is very high there will be a comparatively low number of false positives, meaning a significantly lower overestimation of tumor volume. Recall measures the percentage of true tumor pixels that are classified positively. High recall is especially important for detecting small tumor regions during early diagnosis. Having recoveries with an inverse relation to false positives, the F1-score finds the harmonic means of Precision and Recall, resulting in a single measure of segmentation accuracy. This measure is pertinent when tumor regions are small or imbalanced against the background. An F1, therefore, means that the model detects tumor pixels well with comparatively fewer missed tumor regions, as shown in Equations 33–36.

By optimizing all training hyperparameters via PBFO and splitting the data into training, validation, and test sets, PoSAM-ULTRA achieves robust tumor detection, precise boundary segmentation, and faster convergence, outperforming baseline models on all evaluation metrics.

Proposed study (34) a new network framework called MUNet, which combines the advantages of UNet and Mamba to achieve accurate and efficient segmentation of brain tumors, the results showed that MUNet achieved superior segmentation accuracy and excellent generalization capabilities when validated on TCGA-LGG segmentation datasets, with DSC, IOU, and accuracy reaching 0.702, 0.581, and 0.981, respectively. This paper aims (35) to study the effectiveness of the Segment Anything Model (SAM) in the context of brain MRI, the researchers used the TCGA-LGG dataset, and the results showed that when they applied the SAM model, they were able to achieve 0.87 higher IOU scores.

Researchers (36) aimed to improve the segmentation of low-grade gliomas in MRI by combining a 2D U-Net and a WAO-based voting mechanism. The proposed model achieved an accuracy of 0.9978 and an IOU of 0.8573 on the TCGA-LGG dataset. A hybrid deep learning model combining U-Net and SegNet was proposed for automatic segmentation of low-grade gliomas (LGG) from MRI images (37), and was trained on the TCGA-LGG dataset, where it achieved an average DSC of 83 and 85.7%, outperforming each model individually. Researchers (38) presented an improved U-Net-based segmentation algorithm to improve brain tumor segmentation from MRI images, and after training it on the TCGA-LGG dataset, the proposed method achieved an accuracy of 99.13%, DSC of 0.4596, and IOU of 0.3132.

A lightweight parallel 3D U-Net (LATUP-Net) was designed for brain tumor segmentation from MRI images, ensuring fast and efficient treatment, their proposed model achieved DSC rates of 88.41, 83.82, and 73.67% for the entire tumor, tumor core, and enhanced tumor on the BraTS 2020 dataset, while these rates on the BraTS 2021 dataset were 90.29, 89.54, and 83.92%, respectively (39). In this study (40) the researchers introduced a hybrid model, the 3D ResAttU-Net-Swin, which integrates a residual U-Net, an attention mechanism, and a Swin transformer, the researchers tested their hybrid model on both the BraTS 2020 and BraTS 2019 datasets, the model got a DSC of 88.27% and an IOU of 79.93% on BraTS 2020, and an average DSC of 89.20% and an average IOU of 81.40% on BraTS 2019. In this research paper (41), the VSA-GCNN-BiGRU model was applied to segment and classify brain tumor MRI images to improve diagnostic accuracy, the researchers relied on the BraTS datasets of 2019, 2020, and 2021, the results showed that their proposed model achieved outstanding performance on the BraTS 2019 dataset, with an accuracy of 99.98%, while it achieved an accuracy of 98.2% and a DSC of 40% on the BraTS 2020 dataset, when the proposed model was evaluated on the BraTS 2021 dataset, it achieved an accuracy of 97.6% and a DSC of 98.6%. Table 3 shows a comparison with previous studies on brain tumor segmentation.

The three studies collectively highlight recent advances in deep-learning–based medical image analysis, particularly for improving brain MRI segmentation and classification. The first study (42) introduces a dual-channel encoder–decoder network enhanced with compound attention mechanisms to more accurately segment Multiple Sclerosis (MS) lesions, achieving a notably high Dice score (0.73) and outperforming prior approaches. The second work (43) focuses on brain-tumor classification, leveraging transfer learning with augmented MRI data and demonstrating how pretrained models can significantly boost diagnostic performance when datasets are limited. Meanwhile, the third study (44) presents a supervised segmentation method designed to identify salient brain tissues in MRI images, forming an essential foundation for automated tissue analysis and downstream clinical tasks. The optimized U-Net presented in a recent study (45) achieved remarkably high performance (Dice 92.54%, IoU 90.42%) while maintaining rapid inference time, demonstrating its suitability for real-time clinical applications. Including this reference provides valuable context by highlighting complementary efficiency-focused advancements in deep learning–based medical image segmentation. Together, these studies illustrate a clear progression from precise lesion segmentation to robust tumor classification, to general tissue-segmentation frameworks demonstrating how modern deep learning models continue to enhance diagnostic accuracy and support real-world clinical decision-making.

As illustrated in Table 3, several recent studies on glioma segmentation continue to face recurring limitations that affect their clinical reliability and generalization ability. Many U-Net–based or transformer-enhanced architectures, such as MUNet (34) and the WAO-refined U-Net (36), exhibit overfitting and poor performance in complex or low-contrast tumor regions, particularly when trained on limited single-source datasets. Other methods, including SAM-based segmentation (35), rely heavily on specific evaluation scales and often lack clinical validation or multi-perspective assessment. Several works also suffer from restricted modality usage, as seen in Obayya et al. (38), or do not perform external dataset testing, which limits their generalization capability.

PoSAM-ULTRA is designed to directly tackle these limitations. First, the integration of CBAM, Attention Gates, and multi-scale DownBlocks strengthens tumor-boundary awareness, addressing the poor delineation issues observed in prior CNN-only architectures. Second, the PBFO-driven hyperparameter optimization reduces the tendency toward overfitting by systematically exploring configurations that favor stable, generalizable convergence—unlike models that rely on static or manually tuned settings. Third, the use of a four-channel input, which incorporates prior tumor-likelihood information, enhances robustness across scans with varying intensity distributions, a limitation noted in studies that depend on single MRI modalities.

While we did not conduct an additional external test run in this revision, the comparative evidence in Table 3 demonstrates strong relative generalization: PoSAM-ULTRA surpasses methods known to struggle with dataset shift and maintains performance levels closer to architectures evaluated on heterogeneous clinical datasets (e.g., BraTS 2019–2021). This analysis supports the conclusion that PoSAM-ULTRA not only addresses key limitations in the literature but also exhibits improved resilience when benchmarked against a broad spectrum of contemporary segmentation techniques.

Addressing the need for practical clinical deployment, the PoSAM-ULTRA framework was designed to achieve a favorable balance between high segmentation accuracy and computational efficiency. Our architectural choice of a ResNet-34 encoder offers a strong but relatively lightweight backbone, keeping the overall parameter count manageable. Furthermore, the specialized components CBAM and Attention Gates are implemented to strategically enhance feature discriminative power and optimize information flow across skip connections, which ultimately leads to performance gains without proportional overhead in model size. Critically, the Deep Supervision strategy, integral to stabilizing the training process and accelerating convergence, is disabled entirely during the inference phase. This ensures that PoSAM-ULTRA maintains a high segmentation speed, making it suitable for clinical environments where timely processing is essential, thereby demonstrating a strong commitment to balancing high clinical performance with operational resource efficiency.

The clinical application of PoSAM-ULTRA entails the usage of DICOM protocol support for the integration with hospital PACS systems, thus allowing the automated retrieval and processing of multi-sequence MRI data on GPU-enabled servers. The suggested workflow would be a combination of both automated and manual, wherein, within 1–2 min per case, the system would create preliminary segmentation masks for the radiologist to review and refine through an interactive interface. After the installation, validation on institution-specific data (200–300 cases) is most important to determine the performance across local MRI protocols, along with regulatory compliance (FDA/CE marking), patient consent protocols, and comprehensive staff training. Monitoring segmentation quality and user feedback indisputably assures continuous clinical utility, thereby presenting PoSAM-ULTRA as a smart assistant that not only improves upon diagnostic accuracy but also lightens the radiologist’s workload.

Although PoSAM-ULTRA achieves strong quantitative results, qualitative error inspection remains a crucial component of its clinical usability. During internal validation, the model demonstrated occasional under-segmentation in cases where tumor margins blended gradually with surrounding tissue, making the boundary inherently ambiguous. Similarly, mild over-segmentation was observed in regions exhibiting edema-like textures that share intensity patterns with tumor tissue. Motion-corrupted scans also led to small, localized discrepancies due to artifact-induced distortions in structural information.

These patterns are consistent with the known challenges of glioma segmentation reported in clinical imaging literature. When integrated into the hospital workflow, such errors are efficiently managed through the radiologist-in-the-loop design, where the system generates preliminary masks and the specialist performs rapid verification and refinement. This approach preserves clinical safety while still providing a significant reduction in manual annotation time. Continuous monitoring of these edge cases further supports iterative improvement of the model across diverse imaging protocols.

There are numerous limitations to this study, and they are examiners for future research. The LGG dataset, albeit helpful for the measurement of performance, might not capture the entire variety of the higher-grade tumors and their different imaging protocols in various institutions, thus validation on bigger datasets like BraTS is needed. The present attention on whole tumor segmentation needs to be increased to sub-region delineation, which can contribute to better radiotherapy and surgical planning. Computational power may prove to be a hurdle in the case of resource-limited areas, thus indicating the need for model compression techniques. The model’s low interpretability points out the necessity of merging explainable AI methods and uncertainty quantification to gain clinical trust. Future research will have to include temporal modeling for the longitudinal assessment, multi-modal architectures that combine clinical data (molecular biomarkers, patient history), and extended PBFO optimization for neural architecture search and meta-learning to allow quick adaptation to different healthcare settings.