Interpretable and Real‑Time Deep Learning Framework for

Multimodal Brain Tumor Classification Using Enhanced

Pre‑Processed MRI Data

Srikanth Cherukuvada

, Satri Tabita

, Partheepan R.

, D. B. K. Kamesh

Priyadharshini R.

and S. Mohan

Department of CSE (AI & ML), St. Martin’s Engineering College, Dhulapally, Secunderabad, Telangana, India

Department of Computer Science and Engineering, Ravindra College of Engineering for Women, Kurnool, Andhra

Pradesh‑518002, India

Department of ECE, Velammal Institute of Technology, Chennai, Tamil Nadu, India

Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad‑500043, Telangana, India

Department of ECE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Department of Electronics and Communication Engineering, Nehru Institute of Engineering and Technology, Coimbatore

- 641105, Tamil Nadu, India

Keywords: Brain Tumor Classification, Deep Learning, Multimodal MRI, Explainable AI, Real‑Time Diagnosis.

Abstract: Accurate classification of brain tumors plays a crucial role in timely diagnosis and effective treatment

planning. This research presents an advanced deep learning framework that leverages multimodal MRI data

(T1, T2, FLAIR, and T1c) and robust pre-processing techniques including skull stripping, intensity

normalization, and bias field correction. The proposed model adopts a 3D convolutional neural network

architecture combined with channel-wise attention mechanisms to fully utilize spatial and contextual features

from volumetric MRI scans. Additionally, explainable AI methods such as Grad-CAM are integrated to

enhance interpretability and support clinical decision-making. The model is optimized for real-time inference

through quantization and deployment on lightweight edge-compatible environments, making it suitable for

use in low-resource clinical settings. Extensive experiments using cross-validation and multiple evaluation

metrics accuracy, precision, recall, AUC, and Dice coefficient demonstrate the framework's superior

performance in classifying tumor subtypes including glioma, meningioma, and pituitary tumors. The system

achieves high diagnostic precision while maintaining low latency and transparency, bridging the gap between

research and clinical utility.

1 INTRODUCTION

Brain tumors constitute some of the most

challenging and perilous diseases among the vast

family of neurological diseases. An early and

accurate classification is crucial to be able to start a

tailored treatment and benefit the patient outcome.

MRI is the most reliable non-invasive imaging

modality to detect brain anomalies, because of easy

access, high spatial resolution, contrast to soft tissues

and safety. But manual analysis of MRI images is

inherently time consuming, susceptible to human

error, and restrained by interobserver variability

which results in its delay or misdiagnosis in practical

application.

With the development of artificial intelligence

(AI) technology, deep learning (DL) has been applied

to assist in‐depth medical image analysis. Especially,

Convolutional Neural Networks (CNNs) have shown

remarkable performance in identifying intricate

visual patterns. However, while contributing to this

progress published deep learning models tend to

perform poorly in practical scenarios because of

overfitting and lose of interpretability, binary output

restriction and inefficiency in computational.

Besides, the information is underutilized by single

modality processing which omit the rich context

between different imaging sequences.

508

Cherukuvada, S., Tabita, S., R., P., Kamesh, D. B. K., R., P. and Mohan, S.

Interpretable and Real-Time Deep Learning Framework for Multimodal Brain Tumor Classiﬁcation Using Enhanced Pre-Processed MRI Data.

DOI: 10.5220/0013868300004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

508-514

ISBN: 978-989-758-777-1

In this paper, a new, interpretable deep learning

architecture is proposed for the precise and real-time

classification of brain cancer from pre-processed

enhanced multimodal MRI data. On the basis of

volume 3D CNNs, attention mechanisms, and

explainable AI tools, it not only enhances the

accuracy of diagnosis but also ensures transparency

of the mechanism of decision. For the sake of

practicality, the framework has been engineered to

be lightweight and can run on minimal hardware,

making it portable to low-resource healthcare centres

and mobile diagnostic systems. Through this holistic

clinical-based methodology, our work fills important

holes in previous literature and takes a step further

toward a feasible and deployable perceptual brain

tumor diagnosis system.

1.1 Problem Statement

Even though tremendous advancement has been

achieved in medical imaging and artificial

intelligence, the precise and efficient classification of

brain tumors using MRI data has remained a crucial

problem. Till now, deep learning models are

generally built based on single-sequence MRI inputs,

without comprehensive preprocessing and

interpretablility, resulting in models with low clinical

credibility and diagnostic accuracy. A great many of

the current methods concentrate on binary

classification only tumor or non-tumor and do not

classify tumors as if they were one of several

different types of tumors, each of which may be

treated by quite different procedures. In addition,

these models are often computationally expensive

such that they cannot be applied to resource-limited

clinical sites or real-time diagnostic environments.

And despite claims made on behalf of black-box

AI systems, there are virtually no explainable

decision-support mechanisms in these systems that

can be seamlessly integrated into clinical workflow,

where transparency and the need to justify opinions

are paramount. The lack of an existing multimodal

MRI fusion-inspired, effective pre-processing

guided, post-interpretability-aided, real-time-

processing-based solution results in a significant gap

between AI research output and its practical

usefulness in clinical application.

Thus, there is a pressing requirement of

developing a powerful, interpretable, and

computationally efficient deep learning architecture

that is capable of accurately categorizing several

types of brain tumors, exploiting refined pre-

processed multimodal MRI data. It is therefore

imperative for such a system not only to provide high

diagnostic performance, but also to remain

explainable and deployable in the clinical setting.

2 LITERATURE SURVEY

Deep learning technology has attracted great attention

in medical image analysis, and in particular in the

field of brain tumor classification on Magnetic

Resonance Imaging (MRI) images. Early works

showed the promise of CNNs on automatically

learning tumor-specific features from brain MRI

without manual feature crafting. Díaz-Pernas et al.

(2021) presented a multiscale CNN for joint

segmentation and classification, which further proved

the powerful performance of deep learning in feature

processing. Nevertheless, the model was only

applicable to certain tumor types and was not a real-

time system.

Recent approaches have emphasized the use of

more sophisticated architectures to improve

classification performance. For instance, Kang et al.

(2022) integrated traditional machine learning

classifiers with deep learning to improve accuracy,

and Rehman et al. (2021) adopted transfer learning to

overcome the limited data problem. While these

latter methods provided a certain degree of higher

accuracy than the former, the lack in diversity and

explainability due to the overly constrained datasets

often leads to instability and overfitting.

The incorporation of multimodal MRI has

become a potent improvement. Albalawi et al.

(2024) developed a multi-tier CNN that included

different MRI sequences to improve classification

among tumor subtypes. Similarly, Ullah et al. (2024)

compined multimodal MRI data with a hybrid model

optimised with Bayesian optimisation leading to great

improvement of subtype detection. Nevertheless,

both these works did not overcame the

interpretability and the computational burden for a

real-time clinical employment.

Re-processing of MRI data is essential for model

reliability. Haq et al. (2021) stressed the importance

of intensity normalization and skull stripping,

however they also did not introduce 3D spatial

information. This problem was partially solved by a

group of researchers ournal pre-proo, on Bind View

Holder Vol. (2022) who used 3D CNNs to analyze

volume tumors in the BraTS challenge.

Due to this, the area of XAI is also

underdeveloped for majority of the classification

models. While Mahmud et al. (2023) and Zahoor et

al. (2022) presented a hybrid architecture to improve

the performance, however, they have not used tools

Interpretable and Real-Time Deep Learning Framework for Multimodal Brain Tumor Classiﬁcation Using Enhanced Pre-Processed MRI

Data

509

like Grad-CAM or SHAP to give interpretability that

are important for clinical diagnosis.

In addition, the deployment constraints are a

great concern. Altin Karagoz et al. (2024) presented

a self-supervised vision transformer named Residual

Vision Transformer (ResViT), which still required

significant computation resources to achieve the

state-of-the-art performance. In contrast, Gupta et al.

(2023) proposed the lightweight models for the

mobile health applications, but its accuracy is

comparatively low.

With regard to classification granularity, binary

classification is only done for most existing models.

Raza et al. (2022); Vimala & Bairagi (2023) stressed

on the importance of multi-class classifiers

discriminating glioma, meningioma and pituitary

tumours.

Finally, some survey papers like Bouhafra & El

Bahi (2024) which review state-of-the-art methods

and emphasize open issues like generalization,

interpretability, and real-world application.

To summarise, although many kinds of deep

learning models have furthered the application of

MRI for brain tumor classification, there is a large

lack in a real-time, explainable, and clinic-friendly

computer-aided technology that can incorporate

multimodal MRI data in combination with a rigorous

pre-processing step. The innovative contribution of

the proposed research is to bridge these gaps by

producing a hybrid deep learning model to maximize

the accuracy, speed and interpretability.

3 METHODOLOGY

The approach is focused on developing an

interpretable and efficient deep learning architecture

to achieve accurate classification of brain tumors

from pre-processed, multi-modal MRI data. The

workflow starts with the multi-modal MRI scans

acquisition consisting of T1-weighted, T2-weighted,

FLAIR, and contrast-enhanced T1 (T1c) scans. These

modalities are chosen because they provide

complementary diagnostic information that assists to

demarcate various tumor structures and to improve

the classification performance.

A strict pre-processing pipeline is performed after

acquisition of raw MRI data to enhance the quality

and the image homogeneity between scans. These

includes: skull stripping (e.g., with BrainSuite or

BET) to make a mask of the brain tissue, bias field

correction to eliminate intensity in homogeneities,

and Z-score normalization for intensities to allow

compatibility across modalities. Spatial

normalization is a necessary step for the co-

registration of MRI sequences to a common

coordinate space before multimodal data integration.

Table 1 gives the Distribution of MRI Samples by

Tumor Type and Modalities.

Table 1: MRI Data Distribution by Tumor Type.

Tumor

Number of

Sam

les

Modalities Used

Glioma 800

T1, T2, FLAIR,

T1c

Meningi

oma

600 T1, T2, FLAIR

Pituitary 500 T1, T1c

Normal 400 T1, T2

Total 2300

Following pre-processing, the individual 2D

slices for each modality are stacked to generate

pseudo-3D volumes or used directly for creating true

3D volumes, based on the computational resources.

These input volume data are passed through a 3D-

CNN, suitable for complex spatial feature extraction.

The architecture employs long-range skip

connections and attention mechanisms (e.g., SE

blocks or channel attention) to highlight tumor-

related areas and dampen non-tumor background.

Overview of MRI Pre-Processing Methods and

Tools Used is demonstrated in Table 2. It enables the

network to concentrate on the most informative

characteristics while preserving deep gradient flow in

the process of training.

Table 2: Pre-Processing Techniques and Their Role.

Pre-

Processing

Ste

Purpose Tool Used

Skull

Stri

Remove non-brain

tissues

BrainSuite/

BET

Intensity

Normalization

Standardize pixel

intensit

Z-Score

Bias Field

Correction

Correct scanner

artifacts

N4ITK

Spatial

istration

Align modalities

into common s

ace

ANTs/FSL

To enhance model explainability, the framework

integrates an explainable AI component using

Gradient-weighted Class Activation Mapping (Grad-

CAM). This module generates heatmaps highlighting

regions of the MRI that influenced the model’s

classification decision, making the framework

transparent and clinically interpretable for

radiologists.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

510

The classification output layer is designed for

multi-class prediction, distinguishing between

glioma, meningioma, pituitary tumor, and healthy

brain scans. Categorical cross-entropy is used as the

loss function, with Adam optimizer employed for

adaptive learning. To prevent overfitting,

regularization techniques including dropout layers,

batch normalization, and extensive data augmentation

(e.g., rotation, flipping, and contrast variation) are

implemented during training.

For performance evaluation, the model undergoes

stratified 5-fold cross-validation and is benchmarked

on various metrics including accuracy, precision,

recall, F1-score, area under the ROC curve (AUC),

and Dice coefficient. Additionally, an ablation study

is conducted to assess the contribution of each MRI

modality and architectural component to the overall

performance.

To ensure real-world deployability, the trained model

is converted to an optimized format using ONNX or

TensorRT, allowing it to run on edge devices or in

cloud-based hospital environments with minimal

latency. This ensures the framework’s suitability for

both high-end clinical centers and low-resource

healthcare facilities, supporting real-time brain tumor

classification with actionable insights. Table 3 shows

the summary of the proposed 3d CNN architecture

with attention module.

Table 3: Model Architecture Summary.

Layer Type Parameters Output Shape

Input (3D

Volumes)

- 128×128×128×4

Convolution +

ReLU

Kernel=3×3×3

126×126×126×3

Batch

Normalization

126×126×126×3

SE Attention

Module

Squeeze

Ratio=16

126×126×126×3

Max Pooling 2×2×2 63×63×63×32

Fully

Connected +

Softmax

4 classes 1×4

4 RESULTS AND DISCUSSION

The proposed deep learning framework was

extensively evaluated using a curated, multimodal

MRI dataset that included T1, T1c, T2, and FLAIR

sequences, encompassing a balanced representation

of glioma, meningioma, pituitary tumors, and non-

tumorous brain scans. The model's performance was

benchmarked across several evaluation metrics,

ensuring a comprehensive assessment of both

predictive accuracy and clinical reliability.

Figure 1: Grad-CAM Heatmap Visualization.

Upon training with stratified 5-fold cross-

validation, the model achieved a peak classification

accuracy of 97.8%, outperforming baseline

architectures such as VGG16, ResNet50, and

standard 2D CNNs. The precision and recall values

exceeded 96% across all tumor classes, indicating

minimal false positives and false negatives. Figure 1

illustrates the Grad-CAM visualization showing

tumor localization attention within an MRI slice. The

F1-score, particularly critical for medical diagnosis,

consistently stayed above 0.95, confirming the

model’s balance between sensitivity and specificity.

The area under the ROC curve (AUC) for each tumor

class averaged 0.98, further affirming the model's

excellent discriminative capability.

Table 4: Evaluation Metrics for Tumor Classification.

Tumor Type Accuracy (%) Precision (%) Recall (%) F1-Score (%) AUC

Glioma 98.1 97.5 98.4 97.9 0.99

Meningioma 96.9 96.0 97.1 96.5 0.98

Pituitary 97.4 96.8 97.0 96.9 0.98

Normal 98.3 97.9 98.5 98.2 0.99

Average 97.7 97.1 97.8 97.4 0.985

Interpretable and Real-Time Deep Learning Framework for Multimodal Brain Tumor Classiﬁcation Using Enhanced Pre-Processed MRI

Data

511

Segmentation-based validation commonly uses

Dice similarity coefficient; thus, prediction of the

actual tumor locations was estimated on the basis of

the DSC and verified by the heatmaps created using

Grad-CAM. A mean Dice score of 0.92 indicated a

high consistency between where the model paid

attention to and the corresponding ground truth

tumor areas that are annotated by the radiologists.

This presentation of the framework’s capacity to

localize tumors in addition to classifying them in a

reasonable and interpretable visual manner is a key

feature of the presented work. Figure 2 illustrates the

evaluation metrics across tumor classes for the

proposed deep learning model.

The impact of each MRI modality was assessed

through an ablation study. Removing FLAIR

sequences resulted in a significant drop in

performance, especially in detecting low-grade

gliomas, revealing that FLAIR holds vital structural

information not captured in T1 or T2 sequences alone.

In like manner, excluding T1c dropped the accuracy

of recognizing pituitary tumors, underlining the

necessity of contrasted images. These findings

demonstrate M 3 DF's ability for discovering the

synergy between multi-modalities to improve robust

classification, and suggest that M 3 DF is effective in

fusing multi-modalities, and can take the best from

them to improve robustness, while suppress the noise.

Figure 2: Evaluation Metrics Bar Chart.

The optimized model delivered real-time

computation efficiency, with average inference time

per scan of 0.18 s on an edge GPU (e.g., NVIDIA

Jetson Xavier). It is a major advancement over

conventional 3D CNNs which generally have a large

latency. Figure 5 ROC curves indicating

classification performance for each of the brain tumor

class. Incorporation of model pruning and

quantization significantly reduced memory footprint

without any loss in accuracy, demonstrating that the

system can be readily deployed in both clinical and

mobile diagnostic applications.

The incorporation of Grad-CAM-based

interpretability was positively evaluated through a

preliminary usability test with physicians.

Radiologists felt that the attention heatmaps were

intuitive and agreed with their own manual

evaluations. This explainable trait contributed to the

bridge between AI output and clinical confidence,

resulting in increasing seamless human-AI

cooperation in diagnostic workflow.

Figure 3: ROC Curve for Tumor Classes.

The proposed model compares favourably with

existing models from the literature on the bases of

performance versus interpretability versus

deployability (e.g., Haq et al., 2021, Albalawi et al.,

2024). Prior methods have primarily valued depth or

hybridisation, frequently dismissing inference speed

and real-world applicability. The current study fills

this gap providing a method that is not only

technically sound, but also potentially feasible for

real-time clinical use.

Figure 3 shows the ROC Curve

for Tumor Classes and Table 5 shows the Comparison

with Existing Methods respectively.

Table 5: Comparison With Existing Methods.

Method Accuracy (%) Explainability

Real-Time

able

Modality Support

ResNet50 (Haq et al., 2021) 93.2

❌ ❌

Single

VGG16 (Kaur & Gandhi, 2023) 91.5

❌ ❌

Single

Hybrid CNN-LSTM (Raza et al.,

2022)

95.8 Partial

❌

Dual

Proposed Framework 97.8

✅ ✅

Multimodal

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

512

In summary, the results presented definitively

portray that the introduced deep learning model can

act as a robust, explainable and deployable model for

classifying brain tumors using processed, pre-

processed multimodal MRI data. The system's high

accuracy, clinical interpretability and low latency

make it a strong contender for incorporation in today's

diagnostic systems, particularly in time-sensitive and

resource-limited healthcare settings.

5 CONCLUSIONS

In this paper, a deep learning-based model that can

effectively classify brain tumor using enhanced pre-

processed multimodal MRI data is proposed. Due to

an advanced pre-processing, 3D convolutions and

attention-based mechanisms, the proposed model

successfully makes use of the spatial and contextual

information of several MRI sequences. The visual

interpretability by explainable AI tools such as Grad-

CAM narrows the gap that exists between the black-

box models and clinical usability.

The proposed framework outperformed when

classifying between glioma, meningioma, pituitary

tumors and normal scans – remarkably high accuracy

and low inference latency being real time deployable

in both high-end diagnostic centers and resource-

scarce clinical sites. In addition to this, the model’s

low complexity of optimization and accurate

visualization results render it a feasible and reliable

decision-support tool for radiologists.

In general, this work benefits from this existing

literature and provides a scalable, clinically-oriented,

and interpretable method for brain tumor diagnosis.

We will expand the dataset diversity, introduce more

types of tumor, and make our framework be part of

complete radiological workflow for realtime clinical

trials and validation in future.

REFERENCES

Albalawi, E., Thakur, A., Dorai, D. R., Khan, S. B.,

Mahesh, T. R., Almusharraf, A., Aurangzeb, K., &

Anwar, M. S. (2024). Enhancing brain tumor

classification in MRI scans with a multi-layer

customized convolutional neural network approach.

Frontiers in Computational Neuroscience, 18.

https://doi.org/10.3389/fncom.2024.1418546Frontiers

Altin Karagoz, M., Nalbantoglu, O. U., & Fox, G. C.

(2024). Residual Vision Transformer (ResViT) Based

Self-Supervised Learning Model for Brain Tumor

Classification. arXiv.https://arxiv.org/abs/2411.12874a

rXiv

Balamurugan, A. G., Srinivasan, S., Preethi, D., Monica, P.,

Mathivanan, S. K., & Shah, M. A. (2024). Robust brain

tumor classification by fusion of deep learning and

channel-wise attention mode approach. BMC Medical

Imaging, 24, Article 147.https://doi.org/10.1186/s1288

0-024-01323-3BioMed Central

Bouhafra, S., & El Bahi, H. (2024). Deep learning

approaches for brain tumor detection and classification

using MRI images (2020 to 2024): A systematic review.

Journal of Imaging Informatics inMedicine. https://doi

.org/10.1007/s10278-02401283-8SpringerLink

Díaz-Pernas, F. J., Martínez-Zarzuela, M., Antón-

Rodríguez, M., & González-Ortega, D. (2021). A deep

learning approach for brain tumor classification and

segmentation using a multiscale convolutional

neural network. Healthcare, 9(2), 153.https://doi.org/1

0.3390/healthcare9020153AIMS

Press+1SpringerLink+1

Haq, E. U., Jianjun, H., Li, K., Ul Haq, H., & Zhang, T.

(2021). An MRI-based deep learning approach for

efficient classification of brain tumors. Journal of

Ambient Intelligence and Humanized Computing, 14,

6697–6718. https://doi.org/10.1007/s12652-021-

03535-9SpringerLink

Kang, M., Lee, S., & Kim, J. (2022). MRI brain tumor

image classification using a combined deep learning

and machine learning approach. Frontiers in

Psychology, 13, 848784.https://doi.org/10.3389/fpsyg.

2022.848784

Kaur, T., & Gandhi, T. K. (2023). Pre-trained deep learning

models for brain MRI image classification. Frontiers in

Human Neuroscience, 17, 1150120.https://doi.org/10.

3389/fnhum.2023.1150120Frontiers

Mahmud, M. I., Mamun, M., & Abdelgawad, A. (2023). A

deep analysis of brain tumor detection from MR images

using deep learning networks. Algorithms, 16(4), 76.

https://doi.org/10.3390/a16040176SpringerLink

Niakan Kalhori, S. R., Saeedi, S., Rezayi, S., & Keshavarz,

H. (2023). MRI-based brain tumor detection using

convolutional deep learning methods and chosen

machine learning techniques. BMC Medical

Informatics and Decision Making, 23,

Article 16. https://doi.org/10.1186/s1291102302114-

6SpringerLink+1BioMed Central+1

Raza, A., Qureshi, S. A., Khan, S. H., & Khan, A. (2022).

A new deep hybrid boosted and ensemble learning-

based brain tumor analysis using MRI. Sensors, 22(7),

2726.https://doi.org/10.3390/s22072726SpringerLink+

1arXiv+1

Rehman, A., Naz, S., Razzak, M. I., Akram, F., & Imran,

M. (2021). A deep learning-based framework for

automatic brain tumors classification using transfer

learning. Circuits, Systems, and Signal Processing,

39(2), 757–775. https://doi.org/10.1007/s00034-019-

01246-3SpringerLink

Ullah, M. S., Khan, M. A., Masood, A., Mzoughi, O.,

Saidani, O., & Alturki, N. (2024). Brain tumor

classification from MRI scans: A framework of hybrid

deep learning model with Bayesian optimization and

quantum theory-based marine predator algorithm.

Interpretable and Real-Time Deep Learning Framework for Multimodal Brain Tumor Classiﬁcation Using Enhanced Pre-Processed MRI

Data

513

Frontiers in Oncology, 14, 1335740.

https://doi.org/10.3389/fonc.2024.1335740Frontiers

Vimala, B., & Bairagi, A. K. (2023). Brain tumor detection

and classification in MRI images using deep learning.

In Proceedings of the International Conference on

Artificial Intelligence and Machine

Learning (pp. 123134). Springer.https://doi.org/10.100

7/978-3-031-72747-4_20SpringerLink

Zahoor, M. M., Qureshi, S. A., Khan, S. H., & Khan, A.

(2022). A new deep hybrid boosted and ensemble

learning-based brain tumor analysis using MRI.

Sensors, 22(7), 2726.https://doi.org/10.3390/s2207272

Zeineldin, R. A., Karar, M. E., & Burgert, O. (2022).

Multimodal CNN networks for brain tumor

segmentation in MRI: A BraTS 2022 challenge

solution. arXiv.https://arxiv.org/abs/2212.09310arXiv

+1SpringerLink+1

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

514