ML and DL models have increasingly demonstrated their utility in analyzing clinical images, especially in tasks like segmentation, classification, and anomaly detection. Fatima and Pasha (2017) provided a comparative survey of ML algorithms including Decision Trees (DT), Support Vector Machines (SVM), and K-Nearest Neighbors (KNN), underlining their respective diagnostic strengths and limitations. De Bruijne (2016) traced the application of ML from traditional detection to diagnosis stages in clinical workflows, highlighting the transition from manual to automated decision-making.

Willemink et al. (2020) stressed the importance of robust preprocessing—such as normalization and augmentation—for optimizing ML model performance. Varoquaux and Cheplygina (2022) offered a meta-perspective on methodological challenges and ethical considerations in ML adoption for medical imaging.

Deep learning models, particularly Convolutional Neural Networks (CNNs), have gained prominence due to their high performance in feature extraction and pattern recognition. Zhou et al. (2023) explored various DL models including CNNs, Recurrent Neural Networks (RNNs), and Generative Adversarial Networks (GANs), revealing their effectiveness in multiple imaging modalities like MRI, CT, and histopathology.

Puttagunta and Ravi (2021) showcased the increasing adaptability of DL in tumor detection, while Xu et al. (2014) proposed a hybrid of CNNs and multiple-instance learning to better handle complex feature spaces. Rashed and Popescu (2023) further emphasized DL’s dominance in classification and segmentation tasks.

The convergence of ML and DL has led to the emergence of hybrid models that exploit the strengths of both paradigms. Vadhnani and Singh (2022) surveyed the use of SVM variants with MRI images, noting high accuracy in segmentation and classification tasks. Hussain et al. (2020) developed a hybrid approach incorporating curvelet transformation, ant colony optimization, and SVM to improve image quality and classification accuracy.

Rasool et al. (2022) proposed a hybrid CNN-based architecture to improve tumour classification, while Shahzadi et al. (2018) used a CNN-LSTM model and reported superior performance with features extracted from VGG-16. Khan et al. (2020) employed a comprehensive method using VGG-16/VGG-19 and extreme learning machines (ELM), attaining an accuracy of 92.5%.

Other hybrid efforts include Hashemzehi et al. (2020), who combined CNN with neural autoregressive distribution estimation (NADE) for enhanced classification, and Sun et al. (2019), who used 3D CNNs for both segmentation and survival rate prediction in glioma patients.

For AI models to be clinically acceptable, they must offer transparency in decision-making. Park and Kim (2024) compared various CNN and Transformer architectures, using LIME and SHAP to visualize and explain prediction outputs. Their findings indicate that VGG-16 and ResNet-50, due to their architectural simplicity, produced clearer region-of-interest visualizations than ViT-Base-16.

Narayankar and Baligar (2024) and Mutkule et al. (2023) reviewed several XAI methods such as feature attribution, attention mapping, and rule-based systems. These approaches are gaining traction for their potential to foster clinician trust and regulatory compliance in AI-supported diagnostics.

Despite technical advancements, the deployment of these models in real clinical environments remains limited. Most studies rely on public datasets like BRATS, which lack diversity in patient demographics and imaging protocols (Senan et al., 2022). Additionally, real-time performance, data imbalance, and lack of annotated data limit model robustness (Liu et al., 2021).

Few studies provide comprehensive evaluations of inference time or hardware efficiency, critical for deployment in low-resource settings. Furthermore, regulatory and ethical issues such as data privacy, bias, and explainability remain under-addressed (Alharbi et al., 2023; Aceto et al., 2018).

This study explores the utilization of ML and DL approaches in medical images, particularly in healthcare imaging. One main ML and two DL techniques with Hybrid model are implemented to achieve this goal. We have analysed a range of papers on this subject, examining the techniques proposed and the obstacles encountered when analysing MRI brain tumours using ML, DL, and hybrid machine learning models. Moreover, the study assesses the strengths and shortcomings of the suggested approach to improving the computational efficiency of MRI brain tumour analysis using hybrid machine learning models, which have not been thoroughly examined before. In recent years, numerous studies have employed ML techniques like RNN, ANN, LSTM, SVR, and many more. This study evaluates the improvement in the computational efficiency of MRI brain tumour analysis using hybrid machine learning models, including SVM, VGG-19, YOLOv10, and the SVM + VGG19 Hybrid model.

Data from the full-text selected papers is called Brain MRI Images for Brain Tumour Detection, as its image data. We extracted the following data: journal, publication year, databases searched, study period, setting/scenario, purpose, intervention type, number of studies, study design, main results, opportunities, and implementation challenges. The dataset was obtained from the Kaggle website.

This methodological framework included the subsequent essential elements:

Following this structured protocol, the review sought to deliver a meticulous, organized, and extensive synthesis of the pertinent scientific literature, providing significant insights into the field of study under consideration.

The search process involved querying multiple academic repositories, including Google Scholar, Papers with Code, ScienceDirect, and Springer. The search queries used included:

The initial search retrieved over 923 research articles. A systematic screening process was then applied, where articles were evaluated based on their titles and a brief review of their abstracts. Only studies that effectively addressed the research questions were chosen for further analysis, leading to a final selection of 31 articles; 537 were excluded, as shown in Figure 1.

A thorough selection procedure was implemented to guarantee the pertinence and excellence of the studies included. Papers aligned with the defined research objectives and met the specified criteria were included, while those not following the research scope were excluded. This systematic approach maintained the integrity and formal rigor of the study. The Criteria for Inclusion and Exclusion are as follows:

Table 1 presents the inclusion and exclusion criteria applied during the study selection process. Table 1 provides further clarification on what qualifies as a peer-reviewed source. As indicated in Table 1, only peer-reviewed literature such as journal articles, conference papers, and academic book chapters were considered for inclusion.

Prior to conducting the review, we drafted a written protocol following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (Page et al., 2021; Stovold et al., 2014). The PRISMA statement includes the checklists, explanation and elaboration, and flow diagram.

The PRISMA flow diagram in Figure 1 outlines the study selection process conducted for this systematic review. Initially, 979 records were identified, 923 through database searches and 56 from other registers. Prior to screening, 402 records were removed, including 201 duplicates, 136 records excluded by automation tools, and 65 removed for different reasons. This left 628 records for title and abstract screening, from which 537 were excluded due to irrelevance or failure to meet the inclusion criteria. Of the remaining 72 reports sought for full-text retrieval, 54 could not be retrieved. Consequently, 49 full-text articles were assessed for eligibility. Among these, 31 were excluded as irrelevant, 16 were excluded for not being systematic reviews, and 2 were excluded due to lack of full-text access. Ultimately, 31 studies met the inclusion criteria and were incorporated into the review, represented by 27 individual reports.

Inclusion of quality assessment is a fundamental and critical component of any systematic review process (Salloum, 2018; Salloum et al., 2019; Alhashmi et al., 2019a; Alhashmi et al., 2019b; Alhashmi et al., 2020). This study employed a quality assurance checklist comprising six evaluation questions to assess the methodological rigor of the 31 selected papers, as detailed in Table 2.

Table 2 outlines the quality assurance questions used to evaluate the methodological soundness and clarity of the selected studies. These questions assess key elements such as the clarity of research objectives, the adequacy of methodological explanations, the relevance of findings, and the logical consistency of conclusions. This checklist served as a structured framework to ensure that only studies meeting a minimum standard of academic rigor were included in the review.

Accurate evaluation metrics are crucial for ensuring the effectiveness of MIR brain tumour medical image segmentation in diverse clinical applications. These metrics quantify predicted segments’ similarity and corresponding ground truth annotations. Although the field of MIS has introduced a wide variety of metrics over the past three decades, only a select few have proven to be both appropriate and consistently adopted as standard practices.

Clinical images often exhibit class imbalances, presenting substantial challenges for image segmentation tasks. Standard metrics like specificity or accuracy, which treat true negatives and positives equally, can produce inflated scores even when any pixel is mistakenly identified as the Region of Interest (ROI). This skews evaluation and renders the metrics inappropriate for assessing the effectiveness of segmentation in MIS. Metrics that focus solely on true positive classifications, disregarding true negatives, provide a more accurate assessment for clinical context. Consequently, metrics such as the Dice Similarity Coefficient (DSC) and Intersection-over-Union (IoU) are widely preferred and suggested in MIS (Liu et al., 2021).

Hybrid models, particularly combinations of Support Vector Machines (SVM) with Convolutional Neural Networks (CNNs) such as VGG-19, consistently outperformed traditional and standalone models. These combinations achieved higher accuracy, precision, and recall rates in classification tasks. For instance:

SVM + VGG-19 models exhibited enhanced classification performance, particularly in distinguishing benign from malignant tumours.

YOLOv10n, a lightweight object detection model, achieved near real-time performance with competitive accuracy, making it suitable for resource-constrained clinical settings.

The most common metrics used for performance evaluation included:

Most studies relied on publicly available datasets such as BraTS, Kaggle MRI datasets, and custom institutional collections. However, diversity in patient demographics and imaging protocols was limited, which may affect generalizability.

A classification framework was developed based on the analysis of all 31 research articles included in the systematic review, with each study evaluated in terms of its relevance to the research questions. Papers were marked accordingly when their primary focus aligned with a particular thematic category. For instance, while many articles briefly referenced various applications of brain tumour detection models, only those that provided an in-depth discussion or explicitly concentrated on a particular application were classified under the category segmentation tasks. Figure 2 illustrates the geographical distribution of the reviewed publications. There has been a growing interest in this area over the past two decades, evidenced by the increasing number of publications since 1990, with most contributions originating from the United States.

Hybrid machine learning models, particularly combinations such as SVM with VGG-19 or YOLOv10n, demonstrate significant potential for clinical use in brain tumor diagnosis. These models reduce diagnostic time, minimize human error, and improve detection rates compared to traditional manual interpretations of MRI scans. For instance, the SVM + VGG-19 hybrid achieves high accuracy and precision, making it suitable for classification tasks that can assist radiologists in prioritizing cases with suspected malignancy.

Regarding real-world applicability, models like YOLOv10n are particularly notable due to their lightweight architecture, enabling deployment in resource-constrained environments such as rural clinics or mobile diagnostic units. Their real-time processing capabilities can support faster clinical decision-making and potentially reduce time-to-treatment.

However, despite promising results, significant gaps exist in clinical translation. Most studies utilize retrospective data and control experimental conditions that may not represent the variability and complexity found in real clinical workflows. Challenges include variability in MRI protocols across hospitals, lack of large multi-institutional datasets, integration with existing radiology information systems (RIS), and model explainability, a critical factor for adoption by medical professionals.

Moreover, regulatory and ethical considerations, such as the need for transparency in AI decision-making and the risks of algorithmic bias, remain key barriers. Explainable AI (XAI) techniques like LIME and SHAP can improve trust and interpretability but are still underutilized in current implementations. Therefore, while hybrid models offer high technical performance, successful clinical integration demands a focus on reliability, interpretability, interoperability, and compliance with healthcare regulations.

To demonstrate the competitiveness of the proposed paper, improving the computational efficiency of MRI brain tumour, a comparison is provided in Table 3 showing results from related recent studies whose experiments were conducted using the same or different methods.

The research was pertinent to improving the computational efficiency of MRI brain tumor analysis utilising hybrid ML models. This question highlights its significance and reflects the extensive interest it garners within the field. The consensus was that improving the computational efficiency of MRI brain tumor analysis utilising hybrid ML models were best used for glioma, meningioma, and pituitary tumors, as suggested by Senan et al. (2022), Babu Vimala et al. (2023), and Rashed and Popescu (2023). In addition, hybrid machine learning models can be used for different cancer models (Babu Vimala et al., 2023; Rasool et al., 2022).

Most papers (62.5%) discuss SVM's implementation, limitations, and advantages. Specifically, Senan et al. (2022) compared three different approaches and identified the combination of SVM and the hybrid model as the most promising.

Hybrid models demonstrated superior performance compared to single-method approaches. Integrating traditional ML (e.g., SVM) with DL (e.g., CNNs like VGG-19 or YOLOv10n) enabled:

These benefits are particularly impactful in distinguishing between glioma, meningioma, and pituitary tumours, as reported by Senan et al. (2022) and Rasool et al. (2022).

Hybrid models showed promise in accelerating diagnosis, supporting radiologists, and reducing diagnostic errors. Their use is particularly viable in:

However, clinical implementation remains limited due to gaps in integration with hospital systems and workflow interoperability.

Interpretability remains a significant barrier to clinical acceptance. Few reviewed studies applied Explainable AI (XAI) techniques such as:

Models that incorporated XAI (e.g., VGG-19 + SHAP) produced more transparent decision pathways, allowing radiologists to understand why a tumour was classified as malignant or benign (Park and Kim, 2024).

Despite promising results, several challenges persist:

To enable broader adoption of hybrid models in clinical settings, future work should prioritize:

Beyond performance metrics, the true value of hybrid machine learning models lies in their potential to enhance clinical workflows and support timely, accurate diagnosis of brain tumours. This review shows that these models can significantly reduce computational burden and improve diagnostic performance, but also emphasizes that clinical applicability requires more than just algorithmic success.

Future efforts must prioritize developing models that are accurate and generalizable across diverse populations, compatible with clinical systems, and transparent enough to gain the trust of clinicians. Collaborations with healthcare professionals during the model development process and pilot testing in real clinical environments will be essential to ensure usability, safety, and ethical compliance.

Adopting lightweight, explainable, and clinically validated hybrid models can ultimately contribute to earlier diagnosis, personalized treatment planning, and improved patient outcomes, particularly in under-resourced healthcare settings. As such, hybrid models are not just a technical advancement, but a potential catalyst for more equitable and efficient cancer care.

Future research should prioritize the development of standardized benchmark datasets, the integration of advanced explainable AI (XAI) frameworks such as LIME and SHAP, and the creation of end-to-end pipelines that are both accurate and resource-efficient while maintaining interpretability. In addition, exploring the prospects of hybrid models by incorporating transfer learning and ensemble voting strategies would be highly beneficial. These approaches can enhance model generalizability, robustness, and predictive performance, especially in scenarios with limited annotated medical data.

Moreover, interdisciplinary collaboration between medical professionals and data scientists remains crucial to ensure that developed models meet clinical standards, ethical guidelines, and real-world usability requirements. By focusing on these areas, future work can contribute to advancing intelligent, transparent, and clinically viable solutions for brain tumour detection.