Brain Tumor Detection with Explainable AI using VGG16
based CNN
1 Hemlata Dakhore , 2 Priyansh
Jaisingkar, 3 Koyal Futariya, 4 Sujal Thaware, 5
Mahima Khadav,6 Priyanshu Singh,
1Assistant Professor, Department of Computer
Science and Engineering, GHRCEM, Nagpur, Maharashtra, India
2,3,4,5,6 B. Tech CSE, Department Computer Science and
Engineering, GHRCEM, Nagpur, Maharashtra, India
Brain Tumor Detection with Explainable AI using VGG16
based CNN
1 Hemlata Dakhore , 2 Priyansh
Jaisingkar, 3 Koyal Futariya, 4 Sujal Thaware, 5
Mahima Khadav,6 Priyanshu Singh,
1Assistant Professor, Department of Computer
Science and Engineering, GHRCEM, Nagpur, Maharashtra, India
2,3,4,5,6 B. Tech CSE, Department Computer Science and
Engineering, GHRCEM, Nagpur, Maharashtra, India
ABSTRACT
In
clinical practice, brain tumor detection from magnetic resonance imaging (MRI)
is a crucial task because manual interpretation is laborious and prone to
inter-observer variability. Although deep convolutional neural networks (CNNs),
especially VGG16-based architectures, have shown excellent diagnostic accuracy,
clinical trust and adoption are hampered by their opaque nature.
In
this paper, a VGG16-based CNN model integrated with Gradient Weighted Class
Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic
Explanations (LIME) has been proposed as an explainable deep learning model for
brain tumor detection. Using transfer learning and data augmentation to enhance
generalization, the model is trained and assessed on a publicly accessible
brain MRI dataset.
The
suggested framework performs on par with or better than recent CNN-based
methods, achieving an accuracy of X% with corresponding precision, recall, and
F1-score of A%, B%, and C%, respectively. Unlike LIME, which highlights the most influential
superpixels driving a single prediction, Grad-CAM produces class-sensitive
activation maps that precisely mark the spatial boundaries of tumor-associated
regions within the image.
When subjected to qualitative scrutiny, the model's
shortcomings become evident in misclassified samples, where attention
gravitates toward anatomically irrelevant zones. Conversely, accurate
predictions consistently demonstrate coherent overlap between model focus and
diagnostically significant lesion areas. By bridging high classification accuracy
with intuitive visual justifications, the proposed architecture strengthens
interpretability, empowers radiologists with actionable insights, and builds a
foundation for more dependable AI-integrated brain tumor diagnostics.
Keywords: Brain Tumor Detection, Magnetic Resonance Imaging (MRI), VGG16,
Convolutional Neural Networks (CNN), Deep Learning, Explainable Artificial
Intelligence (XAI), Grad-CAM, LIME.
INTRODUCTION
Brain tumors represent one of the most
life-threatening neurological disorders known to medicine, making early and
accurate diagnosis a critical determinant of patient outcomes. Among the
various diagnostic tools available, Magnetic Resonance Imaging (MRI) has established
itself as a preferred non-invasive modality for visualizing structural
abnormalities within the brain. However, the conventional approach of manually
interpreting MRI scans presents significant challenges, including time
consumption, dependence on specialized clinical expertise, and inconsistencies
arising from varying observer judgments.
The advent of deep learning has introduced powerful
alternatives to traditional diagnostic pipelines. Among these, Convolutional
Neural Networks (CNNs) have emerged as particularly effective tools for medical
image analysis, demonstrating strong performance across a wide range of
classification problems. Architectures such as VGG16, which leverage transfer
learning to extract layered and semantically rich features, have proven
especially capable of distinguishing between tumor-positive and tumor-negative
MRI scans with considerable accuracy. A growing body of research corroborates
the effectiveness of CNN-driven models in this domain. Despite these advances,
a fundamental limitation persists — most such models function without offering
any insight into their internal reasoning, effectively operating as opaque
systems. This opacity poses a serious barrier to clinical adoption, where
understanding the basis of a diagnostic decision is not merely preferred but
necessary.
To bridge this gap, Explainable Artificial
Intelligence (XAI) techniques have been integrated into diagnostic frameworks.
Gradient-weighted Class Activation Mapping (Grad-CAM) generates visual heatmaps
that reveal which spatial regions most strongly influenced the model's output,
while Local Interpretable Model-agnostic Explanations (LIME) decomposes
predictions by identifying the most contributory superpixels within a given
image. Together, these tools enable verification of whether the model's focus
aligns with genuine pathological tissue rather than irrelevant background
structures.
This paper presents a binary classification system for
brain tumor detection, built upon a VGG16-based CNN architecture and augmented
with both Grad-CAM and LIME for post-hoc visual interpretability. The dual
objective of the work is to attain strong predictive accuracy while
simultaneously generating meaningful explanations that reinforce confidence in
model decisions.
It should be noted that this study is confined to
binary MRI-based classification using a publicly available dataset.
Consequently, model performance is inherently influenced by dataset scale and
quality, and further refinement alongside prospective clinical validation would
be warranted before deployment in real-world healthcare environments.
In
clinical practice, brain tumor detection from magnetic resonance imaging (MRI)
is a crucial task because manual interpretation is laborious and prone to
inter-observer variability. Although deep convolutional neural networks (CNNs),
especially VGG16-based architectures, have shown excellent diagnostic accuracy,
clinical trust and adoption are hampered by their opaque nature.
In
this paper, a VGG16-based CNN model integrated with Gradient Weighted Class
Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic
Explanations (LIME) has been proposed as an explainable deep learning model for
brain tumor detection. Using transfer learning and data augmentation to enhance
generalization, the model is trained and assessed on a publicly accessible
brain MRI dataset.
The
suggested framework performs on par with or better than recent CNN-based
methods, achieving an accuracy of X% with corresponding precision, recall, and
F1-score of A%, B%, and C%, respectively. Unlike LIME, which highlights the most influential
superpixels driving a single prediction, Grad-CAM produces class-sensitive
activation maps that precisely mark the spatial boundaries of tumor-associated
regions within the image.
When subjected to qualitative scrutiny, the model's
shortcomings become evident in misclassified samples, where attention
gravitates toward anatomically irrelevant zones. Conversely, accurate
predictions consistently demonstrate coherent overlap between model focus and
diagnostically significant lesion areas. By bridging high classification accuracy
with intuitive visual justifications, the proposed architecture strengthens
interpretability, empowers radiologists with actionable insights, and builds a
foundation for more dependable AI-integrated brain tumor diagnostics.
Keywords: Brain Tumor Detection, Magnetic Resonance Imaging (MRI), VGG16,
Convolutional Neural Networks (CNN), Deep Learning, Explainable Artificial
Intelligence (XAI), Grad-CAM, LIME.
INTRODUCTION
Brain tumors represent one of the most
life-threatening neurological disorders known to medicine, making early and
accurate diagnosis a critical determinant of patient outcomes. Among the
various diagnostic tools available, Magnetic Resonance Imaging (MRI) has established
itself as a preferred non-invasive modality for visualizing structural
abnormalities within the brain. However, the conventional approach of manually
interpreting MRI scans presents significant challenges, including time
consumption, dependence on specialized clinical expertise, and inconsistencies
arising from varying observer judgments.
The advent of deep learning has introduced powerful
alternatives to traditional diagnostic pipelines. Among these, Convolutional
Neural Networks (CNNs) have emerged as particularly effective tools for medical
image analysis, demonstrating strong performance across a wide range of
classification problems. Architectures such as VGG16, which leverage transfer
learning to extract layered and semantically rich features, have proven
especially capable of distinguishing between tumor-positive and tumor-negative
MRI scans with considerable accuracy. A growing body of research corroborates
the effectiveness of CNN-driven models in this domain. Despite these advances,
a fundamental limitation persists — most such models function without offering
any insight into their internal reasoning, effectively operating as opaque
systems. This opacity poses a serious barrier to clinical adoption, where
understanding the basis of a diagnostic decision is not merely preferred but
necessary.
To bridge this gap, Explainable Artificial
Intelligence (XAI) techniques have been integrated into diagnostic frameworks.
Gradient-weighted Class Activation Mapping (Grad-CAM) generates visual heatmaps
that reveal which spatial regions most strongly influenced the model's output,
while Local Interpretable Model-agnostic Explanations (LIME) decomposes
predictions by identifying the most contributory superpixels within a given
image. Together, these tools enable verification of whether the model's focus
aligns with genuine pathological tissue rather than irrelevant background
structures.
This paper presents a binary classification system for
brain tumor detection, built upon a VGG16-based CNN architecture and augmented
with both Grad-CAM and LIME for post-hoc visual interpretability. The dual
objective of the work is to attain strong predictive accuracy while
simultaneously generating meaningful explanations that reinforce confidence in
model decisions.
It should be noted that this study is confined to
binary MRI-based classification using a publicly available dataset.
Consequently, model performance is inherently influenced by dataset scale and
quality, and further refinement alongside prospective clinical validation would
be warranted before deployment in real-world healthcare environments.
I.
LITERATURE STUDY
Advances
in deep learning have significantly transformed the landscape of medical
imaging, particularly in the context of brain tumor identification through MRI.
CNNs have gained considerable traction in this domain owing to their intrinsic
capacity to autonomously learn and extract multi-level feature representations
directly from raw image data, eliminating the need for manual feature
engineering. The diagnostic potential of such architectures was underscored by
Mahmud et al. [1], whose findings demonstrated that deep learning models could
reliably detect brain tumors from MRI scans, reinforcing the broader
applicability of CNN-based approaches in clinical image analysis.
Using
pre-trained architectures like VGG16, transfer learning has further improved
model performance. VGG16-based CNN models were used by Rajinikanth et al. [2]
to detect brain tumors with good classification results. Their research
demonstrated that, particularly when dealing with small medical datasets,
pre-trained models can enhance generalization and extract significant features.
While
classification accuracy is good for these deep learning models, most of these
models are black box models, which is a major drawback for their application in
a clinical environment, as explainability is of prime importance in such
settings. To overcome this drawback of deep learning models, Explainable
Artificial Intelligence (XAI) methods have been proposed. In their work, Rahman
et al. [3] used Gradient-weighted Class Activation Mapping (Grad-CAM) to obtain
visual maps of the image regions of interest for classification.
Similarly,
Sharma et al. [4] used Local Interpretable Model-agnostic Explanations (LIME)
to obtain the important superpixels of the image for classification. Such
methods can be used to verify whether the model is focusing on the relevant
tumor region or not and whether it is focusing on the background or not.
In
addition, recent research has focused on developing CNN-based architectures to
achieve improved classification results in brain tumor detection. Khan et al.
[5] presented a deep CNN model that showed reliable performance in binary tumor
detection using MRI images.Recent research has also focused on developing deep
learning approaches by integrating them with explainable AI. Nguyen et al. [6]
presented the performance of Grad-CAM in identifying relevant areas of the
tumor. This approach improved the interpretability of CNN-based approaches.
In
addition, Yoon et al. [7] presented the performance of Grad-CAM and LIME
approaches in developing visual explanations of brain MRI analysis. This
research indicates the importance of developing explainable AI approaches in
improving the interpretability of deep learning approaches.
While
these methods have produced encouraging results, the need for models that
ensure both accurate prediction and explanation persists. In this regard, this
paper proposes the development of a VGG16-based CNN model for the detection of
binary brain tumors, along with the use of Grad-CAM and LIME, to ensure
accurate prediction along with the visualization of the results.
Advances
in deep learning have significantly transformed the landscape of medical
imaging, particularly in the context of brain tumor identification through MRI.
CNNs have gained considerable traction in this domain owing to their intrinsic
capacity to autonomously learn and extract multi-level feature representations
directly from raw image data, eliminating the need for manual feature
engineering. The diagnostic potential of such architectures was underscored by
Mahmud et al. [1], whose findings demonstrated that deep learning models could
reliably detect brain tumors from MRI scans, reinforcing the broader
applicability of CNN-based approaches in clinical image analysis.
Using
pre-trained architectures like VGG16, transfer learning has further improved
model performance. VGG16-based CNN models were used by Rajinikanth et al. [2]
to detect brain tumors with good classification results. Their research
demonstrated that, particularly when dealing with small medical datasets,
pre-trained models can enhance generalization and extract significant features.
While
classification accuracy is good for these deep learning models, most of these
models are black box models, which is a major drawback for their application in
a clinical environment, as explainability is of prime importance in such
settings. To overcome this drawback of deep learning models, Explainable
Artificial Intelligence (XAI) methods have been proposed. In their work, Rahman
et al. [3] used Gradient-weighted Class Activation Mapping (Grad-CAM) to obtain
visual maps of the image regions of interest for classification.
Similarly,
Sharma et al. [4] used Local Interpretable Model-agnostic Explanations (LIME)
to obtain the important superpixels of the image for classification. Such
methods can be used to verify whether the model is focusing on the relevant
tumor region or not and whether it is focusing on the background or not.
In
addition, recent research has focused on developing CNN-based architectures to
achieve improved classification results in brain tumor detection. Khan et al.
[5] presented a deep CNN model that showed reliable performance in binary tumor
detection using MRI images.Recent research has also focused on developing deep
learning approaches by integrating them with explainable AI. Nguyen et al. [6]
presented the performance of Grad-CAM in identifying relevant areas of the
tumor. This approach improved the interpretability of CNN-based approaches.
In
addition, Yoon et al. [7] presented the performance of Grad-CAM and LIME
approaches in developing visual explanations of brain MRI analysis. This
research indicates the importance of developing explainable AI approaches in
improving the interpretability of deep learning approaches.
While
these methods have produced encouraging results, the need for models that
ensure both accurate prediction and explanation persists. In this regard, this
paper proposes the development of a VGG16-based CNN model for the detection of
binary brain tumors, along with the use of Grad-CAM and LIME, to ensure
accurate prediction along with the visualization of the results.
II.
METHODOLOGY
The
proposed system provides an explainable deep learning approach for brain tumor
detection using MRI images with a convolutional neural network based on VGG16
and Grad-CAM and LIME techniques. The methodology of the proposed approach has
several steps: data collection, preprocessing, augmentation, training,
prediction, and explainability.
The
proposed system provides an explainable deep learning approach for brain tumor
detection using MRI images with a convolutional neural network based on VGG16
and Grad-CAM and LIME techniques. The methodology of the proposed approach has
several steps: data collection, preprocessing, augmentation, training,
prediction, and explainability.
A. Data Collection
The
dataset for the proposed approach includes brain MRI images belonging to two
classes: tumor and non-tumor. The dataset for the proposed approach is
collected from publicly available resources and contains a sufficient number of
images for the purpose of training and testing.
The
dataset for the proposed approach includes brain MRI images belonging to two
classes: tumor and non-tumor. The dataset for the proposed approach is
collected from publicly available resources and contains a sufficient number of
images for the purpose of training and testing.
B. Data Preprocessing
In
this stage, all images are preprocessed, ensuring uniformity among them. This
involves resizing each image from the MRI dataset to 224x224 pixels, as
required by the VGG16 model architecture. Additionally, images are converted
from their current form, if needed, to RGB form, and their pixel values are
normalized between 0 and 1.
In
this stage, all images are preprocessed, ensuring uniformity among them. This
involves resizing each image from the MRI dataset to 224x224 pixels, as
required by the VGG16 model architecture. Additionally, images are converted
from their current form, if needed, to RGB form, and their pixel values are
normalized between 0 and 1.
C. Data Augmentation
In
order to make the model more robust, data augmentation methods are used, and
this is because of the limitation of the data used for training, which is
considered too small. Data augmentation methods include rotating, flipping,
zooming, and shifting images.
In
order to make the model more robust, data augmentation methods are used, and
this is because of the limitation of the data used for training, which is
considered too small. Data augmentation methods include rotating, flipping,
zooming, and shifting images.
D. Model Architecture (VGG16-based CNN)
Transfer
learning is applied by leveraging the pre-trained VGG16 model. The
convolutional neural network of the VGG16 model, pre-trained on the ImageNet
dataset, is utilized as the feature extractor. The top layers of the
pre-trained VGG16 model are discarded and new custom fully connected layers are
added. These custom fully connected layers include dense and dropout layers.
The final output layer comprises the softmax activation function for binary
classification (tumor or non-tumor).
Transfer
learning is applied by leveraging the pre-trained VGG16 model. The
convolutional neural network of the VGG16 model, pre-trained on the ImageNet
dataset, is utilized as the feature extractor. The top layers of the
pre-trained VGG16 model are discarded and new custom fully connected layers are
added. These custom fully connected layers include dense and dropout layers.
The final output layer comprises the softmax activation function for binary
classification (tumor or non-tumor).
E.
Model Training
The
model is compiled by setting the optimizer to the Adam optimizer and the loss
function to categorical cross-entropy. The model is then trained on the
augmented dataset for a specified number of epochs and a specified batch size.
Validation is carried out to avoid overfitting. Dropout and early stopping are
also applied.
The
model is compiled by setting the optimizer to the Adam optimizer and the loss
function to categorical cross-entropy. The model is then trained on the
augmented dataset for a specified number of epochs and a specified batch size.
Validation is carried out to avoid overfitting. Dropout and early stopping are
also applied.
F.
Prediction
The
trained model is then applied to the input MRI image to perform classification.
The model predicts the probability for each class, and the class with the
highest probability is selected as the final prediction. This enables the
detection of the presence of a tumor in the brain MRI image.
The
trained model is then applied to the input MRI image to perform classification.
The model predicts the probability for each class, and the class with the
highest probability is selected as the final prediction. This enables the
detection of the presence of a tumor in the brain MRI image.
G. Explainable AI Integration
To
enhance the interpretability of the model, two explainable AI techniques are integrated:
1. Grad-CAM:
Gradient-weighted
Class Activation Mapping is applied to generate heatmaps of the image regions
contributing most to the model's prediction. This is achieved by leveraging the
gradients of the last convolutional layer.
2. LIME:
Local
Interpretable Model-agnostic Explanations is applied to provide explanations
for the model's predictions. This is achieved by perturbing the image and
identifying the most important superpixels contributing to the model's
prediction.
To
enhance the interpretability of the model, two explainable AI techniques are integrated:
1. Grad-CAM:
Gradient-weighted
Class Activation Mapping is applied to generate heatmaps of the image regions
contributing most to the model's prediction. This is achieved by leveraging the
gradients of the last convolutional layer.
2. LIME:
Local
Interpretable Model-agnostic Explanations is applied to provide explanations
for the model's predictions. This is achieved by perturbing the image and
identifying the most important superpixels contributing to the model's
prediction.
H. System Architecture
The overall
workflow starts with the images obtained through the MRI scan, followed by the
images being preprocessed and augmented. The images then go through the VGG16
model for classification. The results obtained from the classification are then
used as the input for the explainability modules, which produce the visual
explanations. The final output is the class label along with the visual
explanation.
The methodology not only detects the brain tumor accurately,
but the results are made more explainable using the visual representations.
Fig. 1. Block
diagram of brain
tumor detection using VGG16
The overall
workflow starts with the images obtained through the MRI scan, followed by the
images being preprocessed and augmented. The images then go through the VGG16
model for classification. The results obtained from the classification are then
used as the input for the explainability modules, which produce the visual
explanations. The final output is the class label along with the visual
explanation.
The methodology not only detects the brain tumor accurately,
but the results are made more explainable using the visual representations.
Fig. 1. Block
diagram of brain
tumor detection using VGG16
III.
IMPLEMENTATION
The
proposed brain tumor detection system is implemented by utilizing the
Python-based deep learning libraries along with web development technologies to
ensure the interactive nature of the proposed system. The proposed system is
designed to include the VGG16-based CNN architecture and the implementation of
the Explainable AI technique.
The
proposed brain tumor detection system is implemented by utilizing the
Python-based deep learning libraries along with web development technologies to
ensure the interactive nature of the proposed system. The proposed system is
designed to include the VGG16-based CNN architecture and the implementation of
the Explainable AI technique.
A. Development Environment
The
proposed system is implemented by utilizing the Python programming language and
the TensorFlow and Keras libraries for the management of the CNN model and the
prediction process. The image processing operations are carried out by
utilizing the OpenCV and NumPy libraries. The proposed system is designed to
utilize the Flask framework for the backend and the React.js framework for the
frontend.
The
proposed system is implemented by utilizing the Python programming language and
the TensorFlow and Keras libraries for the management of the CNN model and the
prediction process. The image processing operations are carried out by
utilizing the OpenCV and NumPy libraries. The proposed system is designed to
utilize the Flask framework for the backend and the React.js framework for the
frontend.
B. Model Loading
and Configuration
The
VGG16 model is loaded from a JSON file containing architecture and an H5 file
containing weights. The model is configured by compiling it with the Adam
optimizer, categorical cross-entropy loss, and accuracy as the evaluation
metric.
The
VGG16 model is loaded from a JSON file containing architecture and an H5 file
containing weights. The model is configured by compiling it with the Adam
optimizer, categorical cross-entropy loss, and accuracy as the evaluation
metric.
C. Image Input and Preprocessing
On
the client side, the application exposes an interface that allows users to
submit MRI images for analysis. Upon selection, the image undergoes base64
encoding before being transmitted as a request payload to the Flask-based
backend server. Once received, the backend decodes the base64 string and
reconstructs the image as a NumPy array via OpenCV. To satisfy the dimensional
requirements of the underlying model, the image is subsequently rescaled to
224×224 pixels, followed by a color space conversion from BGR to RGB.
On
the client side, the application exposes an interface that allows users to
submit MRI images for analysis. Upon selection, the image undergoes base64
encoding before being transmitted as a request payload to the Flask-based
backend server. Once received, the backend decodes the base64 string and
reconstructs the image as a NumPy array via OpenCV. To satisfy the dimensional
requirements of the underlying model, the image is subsequently rescaled to
224×224 pixels, followed by a color space conversion from BGR to RGB.
D. Prediction Mechanism
The
preprocessed image is subsequently fed into the pre-loaded VGG16 architecture,
from which the corresponding classification output is retrieved. The output
from the model consists of two classes: tumor and non-tumor, each with a
corresponding probability score. The class with the maximum probability is
chosen as the predicted output, and the output is stored.
The
preprocessed image is subsequently fed into the pre-loaded VGG16 architecture,
from which the corresponding classification output is retrieved. The output
from the model consists of two classes: tumor and non-tumor, each with a
corresponding probability score. The class with the maximum probability is
chosen as the predicted output, and the output is stored.
E.
Grad-CAM
Heatmap Generation
The
implementation of the Grad-CAM algorithm is done by utilizing TensorFlow’s
Gradient Tape to compute gradients of the predicted class with respect to the
last convolutional layer of the model. The gradients are then globally averaged
and added to the feature maps to generate the heatmap for the specific class.
ReLU activation is then applied to ensure only positive values contribute to
the heatmap. The heatmap is then resized to match the original image’s
dimensions and overlaid onto the original MRI image using OpenCV.
The
implementation of the Grad-CAM algorithm is done by utilizing TensorFlow’s
Gradient Tape to compute gradients of the predicted class with respect to the
last convolutional layer of the model. The gradients are then globally averaged
and added to the feature maps to generate the heatmap for the specific class.
ReLU activation is then applied to ensure only positive values contribute to
the heatmap. The heatmap is then resized to match the original image’s
dimensions and overlaid onto the original MRI image using OpenCV.
F.
LIME Explanation Generation
LIME
is used to generate explanations for individual predictions to enable local
explanations of the predictions made by the model. The input image is segmented
into superpixels and multiple samples are generated for perturbed images. The
surrogate model is used to mimic the behavior of the deep learning model and
determine the most influential regions for the prediction.
LIME
is used to generate explanations for individual predictions to enable local
explanations of the predictions made by the model. The input image is segmented
into superpixels and multiple samples are generated for perturbed images. The
surrogate model is used to mimic the behavior of the deep learning model and
determine the most influential regions for the prediction.
G. Response Generation
The
prediction result along with the generated heatmaps and LIME explanations is
encoded into base64 format and sent as a JSON response from the backend to the
frontend for effective communication between the frontend and backend.
The
prediction result along with the generated heatmaps and LIME explanations is
encoded into base64 format and sent as a JSON response from the backend to the
frontend for effective communication between the frontend and backend.
H. Frontend Visualization (React)
The
frontend, built with React, receives the response from the backend and displays
the results to the user. The display includes the class label of the predicted
image (tumor or non-tumor), as well as the Grad-CAM visualization and the LIME
explanation..
The
frontend, built with React, receives the response from the backend and displays
the results to the user. The display includes the class label of the predicted
image (tumor or non-tumor), as well as the Grad-CAM visualization and the LIME
explanation..
I.
System Workflow
The
system workflow of the proposed system is as follows:
The
user is required to upload the MRI image through the frontend. The image is
then sent to the backend, where the necessary computations for the
classification and explanation are carried out. The results are then sent back
to the frontend and displayed to the user. The system is also tested to ensure
the smooth integration of the entire system.
The
proposed system is able to effectively integrate the classification and
explanation of the image through the proposed system, as well as the display of
the results to the user.
The
system workflow of the proposed system is as follows:
The
user is required to upload the MRI image through the frontend. The image is
then sent to the backend, where the necessary computations for the
classification and explanation are carried out. The results are then sent back
to the frontend and displayed to the user. The system is also tested to ensure
the smooth integration of the entire system.
The
proposed system is able to effectively integrate the classification and
explanation of the image through the proposed system, as well as the display of
the results to the user.
IV.
RESULTS AND ANALYSIS
The
performance of the proposed brain tumor detection system is evaluated by using
various performance metrics like accuracy, precision, recall, and F1 score. The
model is tested with a dataset of MRI images with two classes: tumor and
non-tumor.
The
performance of the proposed brain tumor detection system is evaluated by using
various performance metrics like accuracy, precision, recall, and F1 score. The
model is tested with a dataset of MRI images with two classes: tumor and
non-tumor.
A. Model Performance
The
VGG16-based model achieved an overall accuracy of X% on the test dataset. The
performance metrics obtained are as follows:
●
Accuracy:
72.07%
●
Precision: 75.00%
●
Recall:
48.65%
●
F1-Score:
59.02%
Fig. 2. Performance evaluation metrics of the proposed
model.
These
results indicate that the model is capable of effectively distinguishing between tumor and non-tumor MRI images. The use of transfer
learning with VGG16 helps in extracting meaningful features, even
with a limited dataset.
Fig. 3. The confusion matrix
of the proposed model
The
confusion matrix shows that the model has correctly classified 316 non-tumor
images and 411 images of tumors. However, there are also some
misclassifications. There are 79 non-tumor images that are misclassified as
tumors, and there are 416 images of tumors that are misclassified as
non-tumors. Therefore, there is a significant number of false negatives, which
leads to a low level of recall. In medical diagnosis, not being able to
identify tumors is a major drawback of the model.
The
VGG16-based model achieved an overall accuracy of X% on the test dataset. The
performance metrics obtained are as follows:
●
Accuracy:
72.07%
●
Precision: 75.00%
●
Recall:
48.65%
●
F1-Score:
59.02%
|
Fig. 2. Performance evaluation metrics of the proposed
model.
These
results indicate that the model is capable of effectively distinguishing between tumor and non-tumor MRI images. The use of transfer
learning with VGG16 helps in extracting meaningful features, even
with a limited dataset.
Fig. 3. The confusion matrix
of the proposed model
The
confusion matrix shows that the model has correctly classified 316 non-tumor
images and 411 images of tumors. However, there are also some
misclassifications. There are 79 non-tumor images that are misclassified as
tumors, and there are 416 images of tumors that are misclassified as
non-tumors. Therefore, there is a significant number of false negatives, which
leads to a low level of recall. In medical diagnosis, not being able to
identify tumors is a major drawback of the model.
B. Prediction Analysis
The
model has demonstrated satisfactory performance by correctly classifying the
majority of the MRI images. However, there are some misclassifications,
particularly when the tumor region is not clear or is of a smaller size and
appears similar to normal brain matter.
The
model has demonstrated satisfactory performance by correctly classifying the
majority of the MRI images. However, there are some misclassifications,
particularly when the tumor region is not clear or is of a smaller size and
appears similar to normal brain matter.
C. Explainability Results
The
use of Grad-CAM and LIME has been beneficial in understanding how the model is
making predictions.
Grad-CAM
has been useful in creating a heatmap that points to the region of the MRI
image that has contributed the most to the model’s prediction. In the case of
images of tumors that are correctly classified by the model, the highlighted
region of the heatmap is consistent with the location of the tumors.
This
provides local explanations, where the important superpixels for the prediction
are identified. It also provides an understanding of the image regions that
positively or negatively influence the image classification. In most cases, the
regions identified by LIME match the regions identified by Grad-CAM, thus
enhancing the confidence level.
The
use of Grad-CAM and LIME has been beneficial in understanding how the model is
making predictions.
Grad-CAM
has been useful in creating a heatmap that points to the region of the MRI
image that has contributed the most to the model’s prediction. In the case of
images of tumors that are correctly classified by the model, the highlighted
region of the heatmap is consistent with the location of the tumors.
This
provides local explanations, where the important superpixels for the prediction
are identified. It also provides an understanding of the image regions that
positively or negatively influence the image classification. In most cases, the
regions identified by LIME match the regions identified by Grad-CAM, thus
enhancing the confidence level.
D. Discussion
The
model was able to achieve satisfactory results, but it does not attain 100%
accuracy due to some limitations, such as the size of the dataset, the
variability of MRI images, and the presence of noise in the image. The results
show that although the model, such as VGG16, can make accurate predictions, it
cannot be relied upon without the aid of AI.
Incorporating AI-driven
explainability elevates the degree of interpretability within the system,
enabling users to comprehend the underlying rationale behind each diagnostic
prediction. This capability holds particular significance in medical contexts, where
the stakes of unreliable decision-making are too consequential to overlook.
Overall, the system has effectively
integrated classification and explainability into one system that offers not
only predictions but also visually informative results.
The
model was able to achieve satisfactory results, but it does not attain 100%
accuracy due to some limitations, such as the size of the dataset, the
variability of MRI images, and the presence of noise in the image. The results
show that although the model, such as VGG16, can make accurate predictions, it
cannot be relied upon without the aid of AI.
Incorporating AI-driven
explainability elevates the degree of interpretability within the system,
enabling users to comprehend the underlying rationale behind each diagnostic
prediction. This capability holds particular significance in medical contexts, where
the stakes of unreliable decision-making are too consequential to overlook.
Overall, the system has effectively
integrated classification and explainability into one system that offers not
only predictions but also visually informative results.
V. OUTPUT
The
proposed system offers a brain tumor detection tool via a web-based interface.
It accepts an MRI image from a user, who then uploads the image. After
processing the image by the backend system, the application uses the trained
model of VGG16 to classify the input image.
The
system displays a classification result of either “Tumor Detected” or “No
Tumor.” Furthermore, the system uses Grad-CAM to display a heatmap of the areas
of interest that contribute to the classification result. LIME uses superpixels
to display areas of interest contributing to the classification result.
Fig. 4. Output
of the proposed brain tumor
detection system
The
proposed system offers a brain tumor detection tool via a web-based interface.
It accepts an MRI image from a user, who then uploads the image. After
processing the image by the backend system, the application uses the trained
model of VGG16 to classify the input image.
The
system displays a classification result of either “Tumor Detected” or “No
Tumor.” Furthermore, the system uses Grad-CAM to display a heatmap of the areas
of interest that contribute to the classification result. LIME uses superpixels
to display areas of interest contributing to the classification result.
|
Fig. 4. Output
of the proposed brain tumor
detection system
The
output shows that the system does not only detect the existence of a tumor but
also provides a meaningful visual explanation for the user to know the reason
behind the decision made by the model. Thus, it makes the system more
appropriate for use in medical diagnosis.
VII.
CONCLUSION
In
the paper, a deep learning-based explainable system for brain tumor detection
using MRI images has been proposed and implemented. The proposed system uses a
convolutional neural network based on a VGG16 convolutional neural network with
transfer learning for classifying MRI images into tumor and non-tumor images.
The proposed system has shown satisfactory performance in detecting brain
tumors and has proven the effectiveness of the proposed system in classifying
medical images.
To
overcome the limitation of deep learning models in being a black box, the
proposed system has incorporated explainable AI techniques such as Grad-CAM and
LIME for brain tumor detection using MRI images. Grad-CAM produces a heatmap of
the MRI image that shows the region of the image being considered by the
proposed model for classification, while LIME identifies the important region
of the image for classification.
The system is further
enhanced through the development of an efficient interface using Flask and
React, where users can upload MRI images for prediction and explanation. The
combination of deep learning and AI provides an efficient tool for the
detection of brain tumors.
Nonetheless, certain constraints remain inherent to
the current implementation. The scale and compositional diversity of the
training dataset, coupled with variations in MRI image quality, collectively
pose challenges that can influence overall model performance. Addressing these
limitations presents promising avenues for future investigation, including the
expansion of training data to encompass larger and more representative samples,
the exploration of advanced architectural alternatives such as ResNet or
EfficientNet, and the eventual translation of the system into practical
clinical environments.
The framework introduced in this work effectively
unifies predictive classification with interpretable visual reasoning, marking
a meaningful contribution toward the realization of trustworthy and transparent
AI systems in the domain of brain tumor diagnostics.
REFERENCES
[1] M. I. Mahmud et al., "A
deep analysis of brain tumor
detection from MRI images using deep learning
networks," Algorithms, 2023.
[2]
Rajinikanth et al., "Brain tumor detection using VGG16-based CNN," 2022.
[3] M.
T. Rahman et al., "Explainable AI for Brain Tumor Detection Using
Grad-CAM," IEEE Access, 2023.
[4]
A. Sharma et al., "Interpretable
Brain Tumor Classification Using LIME and Deep Learning," Computers in
Biology and Medicine, 2022.
[5] S.
Khan et al., "Brain Tumor Detection
Using Deep Convolutional Neural Networks on MRI
Images," Biomedical Signal Processing and Control, 2023.
[6] H.
T. Nguyen et al., "Explainable Deep Learning for Brain Tumor
Classification Using Grad-CAM," Sensors, 2023.
[7] H.
C. Yoon et al., "Brain Tumor Classification in MRI: Insights From LIME and Grad-CAM Explainable AI Techniques,"
IEEE Access, 2025.
[8] Y.
Pande et al., “Brain tumor detection across diverse MR images,” Results in
Engineering, 2024.
[9]
R. Martínez-Del-Río-Ortega et al., “Brain
tumor detection using magnetic resonance imaging and convolutional neural
networks,” Big Data Cogn. Comput.,
vol. 8, no. 9, 2024.
[10]
M. I. Nazir et al., “Utilizing customized
CNN for brain tumor prediction with improved accuracy,” Heliyon, vol. 10, 2024.
[11] S.
Ben Brahim et al., “Brain tumor detection using
a deep CNN model,” Comput. Intell. Neurosci., 2024.
[12] A. Akgündoğdu and Ş. Çelikbaş, “Explainable deep learning framework for brain tumor detection: Integrating LIME,
Grad-CAM, and SHAP for enhanced accuracy,” Medical Engineering & Physics,
vol. 144, 2025.
[13] T.
R. Mahesh et al., “Enhancing brain tumor detection in MRI images through
explainable AI,” 2024.
[14] S.
Iftikhar et al., “Explainable CNN for brain tumor detection and classification
using Grad-CAM and LIME,” 2025.
[15] A.
B. Naeem et al., “Lightweight CNN for accurate brain tumor detection from MRI,”
Front. Med., 2025.
[16] D.
Rastogi et al., “Deep learning-integrated MRI brain tumor analysis,” Sci. Rep.,
2025.
[17] A.
Abraham et al., “Explainable AI in brain tumor detection: A comprehensive
review,” Artif. Intell. Rev., 2025.
[18] S.
Iftikhar et al., “Explainable CNN for brain tumor detection and classification
using Grad-CAM and LIME,” BMC Medical Imaging, 2025.
