Layer-Wise Relevance Propagation for Classifying Brain MRI Images

Ganesh Naik, Shivyogi Bendegerimath, Vijeth Kawari,

Gautam Narajji and Prashant Narayankar

School of Computer Science and Engineering, KLE Technological University, Hubli, Karnataka, India

Keywords:

Brain Tumor Classiﬁcation, Explainable AI (XAI), Medical Imaging, Layer-Wise Relevance Propagation

(LRP), Automated Medical Report Generation.

Abstract:

Accurate diagnosis and explainable predictions are important in effective planning and monitoring treatment in

brain tumor analysis using medical imaging. To enhance the capabilities of tumor detection and interpretation

in Brain MRI scans, the proposed work presents a comprehensive framework that combines Explainable AI

(XAI) with brain tumor classiﬁcation. The framework, based on ResNet18, a deep learning model, classiﬁes

MRI images into four categories: glioma tumor, meningioma tumor, pituitary tumor, and no tumor. The

system incorporates Layer-wise Relevance Propagation (LRP) to highlight regions inﬂuencing predictions,

providing richer interpretability and visual explanations of the decision-making process. The proposed work

has demonstration of 3 approaches for explainable decision making process 1) LRP with heatmaps 2) LRP

using overlayed heatmaps 3) Pixel-wise Relevance of presence of tumor. Additionally, the proposed approach

includes Automated Medical Report Generation, summarizing categorization results and presenting visual

explanations to assist physicians effectively. The proposed model has reached 85% accuracy with strong

prediction capabilities and superior explainability in performance to adequately fulﬁll the fundamental demand

of AI-based health solutions to provide more transparent and reliable performance.

1 INTRODUCTION

Brain tumors are among the most serious and danger-

ous types of cancer; they disturb the delicate balance

of the brain and impair vital physical and cognitive

functions. The diagnosis is especially challenging be-

cause of the complex anatomy of the brain and be-

cause the tumors have a tendency to simulate healthy

tissue in imaging studies (Litjens et al., 2017a).

In Magnetic Resonance Imaging (MRI), despite ad-

vancements, diagnoses remain time-consuming and

heavily reliant on human judgment. This reliance

delays critical treatments, particularly for aggressive

cancers. Chemotherapy, radiation, and surgery also

carry signiﬁcant risks, making surgical precision cru-

cial to avoid damaging surrounding tissues. These

challenges prolong the emotional and mental stress

experienced by patients and their families, emphasiz-

ing the need for timely and precise diagnostic solu-

tions.

Conventional MRI analysis requires interpretation

by radiologists, which is subject to human error and

delays the early diagnosis and treatment of brain tu-

mors (Ronneberger et al., 2015). Furthermore, many

existing AI algorithms function as “black boxes,” pro-

viding predictions without explaining their rationale.

This lack of transparency hampers medical profes-

sionals, who require clear, interpretable information

to make informed decisions. Addressing these limita-

tions is critical to advancing diagnostic accuracy and

building trust in AI-powered tools.

The proposed work aims to address these chal-

lenges using advanced machine learning techniques,

such as Deep Learning and Layer-wise Relevance

Propagation (LRP) (Bach et al., 2015a). By lever-

aging the ResNet architecture, the model effectively

processes high-dimensional MRI datasets and han-

dles complex analysis tasks, such as identifying ir-

regular shapes and varying intensities in tumors (He

et al., 2016). The integration of LRP provides inter-

pretable heatmaps that highlight critical MRI regions

inﬂuencing classiﬁcation decisions, fostering trans-

parency and trust among medical professionals. The

approach further reduces dependence on traditional

manual feature engineering, making it suitable for

low-resource settings where expert radiologists may

be scarce. Additionally, the system supports accu-

rate tumor subtype characterization and visual expla-

Naik, G., Bendegerimath, S., Kawari, V., Narajji, G. and Narayankar, P.

Layer-Wise Relevance Propagation for Classifying Brain MRI Images.

DOI: 10.5220/0013607800004664

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

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 5-11

ISBN: 978-989-758-763-4

nations of tumor-speciﬁc properties, such as hetero-

geneity and border irregularity, enabling personalized

therapeutic interventions (Isensee et al., 2021).

Brain tumors affect an estimated 40,000 to 50,000

adults annually in India, with children constituting

20% of these cases. This prevalence, combined with

the unique challenges posed by brain tumors, empha-

sizes the urgent need for improved diagnostic tools.

Many existing algorithms remain opaque, operating

as “black boxes” and providing predictions without

explaining the rationale. The lack of transparency

hampers medical professionals, who rely on clear and

precise information for making informed decisions.

Addressing this gap is crucial to advancing patient

care and fostering conﬁdence in AI-powered solu-

tions.

The rest of the paper is organized as follows:

Section II reviews the relevant literature and high-

lights existing gaps in the domain. Section III de-

tails the methodology, including the ResNet architec-

ture, LRP integration along with Report Generation.

Section IV presents experimental results and analysis,

showcasing the model’s effectiveness in addressing

diagnostic challenges. Section V discusses the clin-

ical applicability of the proposed approach and its po-

tential impact along with future research directions.

2 BACKGROUND STUDY

2.1 Related Work and Prior Studies

Recent advancements in medical imaging have fo-

cused on multi-class classiﬁcation of brain MRI im-

ages, with deep learning models achieving signiﬁcant

breakthroughs in accuracy and efﬁciency. Traditional

machine learning techniques, such as Support Vector

Machines (SVMs) and K-Nearest Neighbors (KNNs),

relied heavily on handcrafted features like Gray-

Level Co-occurrence Matrices (GLCM) and Princi-

pal Component Analysis (PCA) (Bach et al., 2015b).

These methods often struggled with the complexity

and variability inherent in medical imaging datasets.

In contrast, Convolutional Neural Networks (CNNs)

and transfer learning frameworks, including ResNet,

AlexNet, and GoogLeNet, have demonstrated supe-

rior robustness and scalability in classifying MRI im-

ages into multiple classes (Vankdothu and Hameed,

2022). The incorporation of preprocessing techniques

such as data augmentation, skull stripping, and mor-

phological operations further enhances the effective-

ness of these models, showcasing their potential for

clinical applications (Kulkarni and Sundari, 2020).

Despite their success in achieving high classiﬁ-

cation accuracy, deep learning models often suffer

from a black-box nature, which hinders their inter-

pretability and transparency in medical imaging. Ex-

plainability is critical in multi-class classiﬁcation, as

understanding the reasoning behind predictions fos-

ters reliability and trust among clinicians. Visu-

alization methods like Grad-CAM have been em-

ployed to highlight tumor regions in MRI images,

adding a layer of interpretability to these models

(Pang et al., 2023). Additionally, techniques such

as Layer-wise Relevance Propagation (LRP) have

emerged as powerful tools for explaining classiﬁer

decisions by providing pixel-wise decomposition of

predictions. LRP generates heatmaps that highlight

regions most relevant to a given class prediction, en-

hancing transparency and interpretability (Bach et al.,

2015b). Studies have validated LRP’s utility in multi-

class medical imaging tasks by conﬁrming predic-

tions and identifying biologically meaningful fea-

tures, reinforcing its value in AI-driven diagnostic

systems (Babu Vimala et al., ).

2.2 Gaps in Current Research and How

proposed work Addresses Them

While existing deep learning models have achieved

remarkable performance in classifying brain MRI im-

ages, they often lack adequate interpretability. Meth-

ods like Grad-CAM, though widely used, focus pri-

marily on high-level feature activations and lack the

precision required for ﬁne-grained analysis. Addi-

tionally, they may fail to distinguish subtle differ-

ences among multiple classes, a critical need in med-

ical imaging (Pang et al., 2023). Furthermore, tech-

niques like SHAP (SHapley Additive exPlanations),

which emphasize feature importance, are computa-

tionally expensive and do not provide the spatial vi-

sualizations necessary for medical diagnostics.

These limitations highlight the need for more ad-

vanced explainability methods, such as Layer-wise

Relevance Propagation (LRP), which combines com-

putational efﬁciency with detailed interpretability.

LRP addresses the black-box challenge by generat-

ing pixel-wise heatmaps that pinpoint the regions con-

tributing most to model predictions, providing a ﬁne-

grained understanding of decision-making processes

(Bach et al., 2015b). Unlike Grad-CAM, LRP en-

sures granularity in analyzing multi-class predictions,

making it suitable for distinguishing subtle differ-

ences in tumor characteristics. Additionally, the pro-

posed integration of LRP with advanced architectures

like ResNet leverages the strengths of deep learn-

ing for robust multi-class classiﬁcation while enhanc-

ing transparency. By addressing these gaps, the cur-

INCOFT 2025 - International Conference on Futuristic Technology

rent work bridges the divide between high-performing

deep learning models and the clinical need for ex-

plainable AI solutions, promoting trust and adoption

among medical professionals.

3 PROPOSED WORK

The proposed model for brain tumor classiﬁcation

uses ResNet-18 while for explaining the decision-

making process, it uses Layer-wise Relevance

Propagation, a technique used in Explainable Artiﬁ-

cial Intelligence (XAI) to make deep learning models,

especially neural networks, more interpretable. For

making the classiﬁcation and explanation process

more readable, the model uses a report generation

module that summarizes the intricate details of

tumor detection and analysis. The workﬂow for the

proposed model is shown in Figure 1.

Figure 1: Proposed Model Workﬂow.

The following steps make up the methodology:

3.1 Dataset Preparation

Dataset consists of training and testing folders of

Brain MRI images of namely 4 classes: glioma tumor,

pituitary tumor, meningioma tumor and no tumor.

During the Dataset Preparation process, photos

are categorized by their class names and arranged

into distinct training and testing folders. The image

is shrunk to 224 × 224 and normalized using stan-

dard ImageNet mean and standard deviation data. Py-

Torch’s ImageFolder tool is used to load the dataset,

and data loaders are made with a batch size of 32 for

processing efﬁciency.

The formula for normalization is shown in Equation

normalized

x −µ

, (1)

3.2 Resnet-18

Resnet-18 architecture as shown in Figure 2 is a deep

learning architecture that is very powerful in handling

complex image classiﬁcation tasks. The residual con-

nections make ResNet-18 particularly effective, as

they allow the network to learn deeper and more intri-

cate features without running into problems like van-

ishing gradients. These connections enable the model

to process and understand detailed patterns in images,

which is crucial when working with medical images

like MRIs. ResNet-18 has been used in a number of

image recognition applications, and its ability to cap-

ture simple features such as edges and more complex

patterns that differentiate tumor types makes it a good

choice for this task. Besides, ResNet-18 is computa-

tionally efﬁcient, and it can be ﬁne-tuned for speciﬁc

tasks like brain tumor classiﬁcation without requiring

much more resources. This makes ResNet-18 a good

model for this classiﬁcation task due to its accuracy,

efﬁciency, and ability to handle complex data.

The fundamental architecture we use is the

ResNet-18 model pretrained on the ImageNet dataset.

The fully connected layer is swapped out for a

new linear layer that generates probabilities for four

classes in order to modify the model for brain tumor

classiﬁcation. The brain tumor dataset is then used to

reﬁne the model.

The formula for the output of the linear layer is shown

in Equation 2:

y = softmax(Wh + b) (2)

Figure 2: Resnet18 Architecture for Brain Tumor Classiﬁ-

cation.

3.3 Training and Optimization

In Training and Optimization, the Adam optimizer

with a learning rate of 0.001 is used to optimize the

model parameters during the training phase. The error

between the true and anticipated class labels is calcu-

lated using the cross-entropy loss function. Over the

Layer-Wise Relevance Propagation for Classifying Brain MRI Images

course of 50 epochs, the training loop iteratively mod-

iﬁes the model’s parameters while tracking accuracy

and loss, among other performance metrics, through-

out training and validation.

The formula for Cross-Entropy Loss is shown in

Equation 3:

L = −

∑

i=1

∑

j=1

i j

log( ˆy

i j

) (3)

The formula for Adam Optimizer is shown in

Equation 4:

= θ

t−1

−η

√

+ ε

(4)

3.4 Predictability and Explainability

The model uses the training weights to predict the

class of a particular MRI picture in this step of pre-

dictability and explainability. Layer-wise Relevance

Propagation (LRP) is used to make the data inter-

pretable. By breaking down the model’s predictions

into pixel-by-pixel relevance scores, LRP highlights

the areas of the MRI that have the most inﬂuence on

the outcome.

The formula for relevance propagation is shown in

Equation 5:

∑

i j

∑

k j

(5)

3.5 Performance Evaluation

Lastly, accuracy is the primary metric used to measure

the performance of the model during performance

evaluation. Both training and validation accuracy are

monitored throughout the training process to ensure

robustness and generalization.

The formula for accuracy is given by Equation 6:

Accuracy =

Number of correct predictions

Total number of predictions

(6)

3.6 Report Generation

A comprehensive PDF report is generated at the end

of the model evaluation process.The report provides a

summary of the visualizations, explanations, and key

ﬁndings in the experiment. It includes patient-speciﬁc

information such as whether or not a tumor was de-

tected, the type of tumor, and the region highlighted

by the heatmap corresponding to the tumor. More-

over, the report includes medical advice for the patient

according to the type of tumor; it might include sug-

gestions for periodic monitoring or potential surgical

removal. It also serves as a useful tool for clinicians

and researchers, providing not only visual outputs but

also contextual information to be used in the decision-

making and further analysis.

4 RESULTS AND ANALYSIS

In this section, we present and analyze the results

of the proposed model, which demonstrated high ac-

curacy in classifying brain tumors into glioma tu-

mor, meningioma tumor, pituitary tumor, and non-

tumorous cases. The performance metrics, such as

precision and F1-score, validate the model’s ability to

perform the task effectively. Additionally, Layer-wise

Relevance Propagation (LRP) was utilized to enhance

interpretability, providing heatmaps that highlight the

location of the tumor.

4.1 Presentation of Results through

Visualizations

To provide clear and effective insights into the

model’s performance, we use three different types of

visualizations based on LRP. These visualizations not

only demonstrate the model’s focus but also help as-

sess its interpretability.

The following results are for the type of Meningioma

tumor:

Figure 3: Layer-wise Relevance Propagation with

heatmaps.

The heatmap shown in Figure 3 highlights areas of

high relevance using warm colors, where the darkest

red indicates the strongest attention by the model. The

areas that correspond to the tumor region are clearly

visible, making it evident that the model is focusing

on clinically signiﬁcant structures.

Figure 4 shows the overlayed heatmap, which

combines the heatmap with the original MRI image.

This allows us to visualize the anatomical structure

along with the relevance information. This technique

INCOFT 2025 - International Conference on Futuristic Technology

Figure 4: Layer-wise Relevance Propagation Using Over-

layed Heatmaps.

provides a spatial context for the model’s focus, en-

suring that the relevance aligns with clinical expec-

tations and helping users understand the relationship

between the tumor and surrounding tissues.

Figure 5: Pixel-wise Relevance of Presence of Tumor.

Figure 5 presents LRP without the heatmap

overlay, offering a simpler depiction of the relevant

areas. This visualization provides a straightforward

representation of relevance, using binary masks to

identify tumor regions, as well as showing non-

contributing areas and tumor contours with distinct

color coding: blue for tumor regions, red for tumor

borders, and green for non-relevant areas.

Similarly the relevance maps for pituitary ( Figure

6, Figure 7) and glioma ( Figure 8, Figure 9) type of

tumors are also depicted. The model does not gener-

ate any type of map if tumor is not detected.

Figure 6: Pixel-wise Relevance of Pituitary Tumor

Figure 7: Overlayed heatmap of Pituitary Tumor

Figure 8: Pixel-wise Relevance of Glioma Tumor

Figure 9: Overlayed heatmap of Glioma Tumor

4.2 Detailed Analysis of Model

Performance

The model demonstrated high accuracy in identifying

and classifying brain tumors. Performance metrics,

including precision and F1-score, validate its effec-

tiveness. The use of LRP heatmaps further enhances

the conﬁdence in the model’s decision-making pro-

cess by visually conﬁrming that it focuses on the cor-

rect areas, such as the tumor regions.

The heatmaps provide an intuitive understanding

of the model’s attention to speciﬁc areas, while the

overlayed heatmaps give a more contextual view, al-

lowing clinicians to assess the spatial relationship be-

tween the tumor and its surrounding tissue. These vi-

sualizations not only validate the model’s predictions

but also serve as a reliable interpretability tool that

can support clinical decision-making.

Layer-Wise Relevance Propagation for Classifying Brain MRI Images

4.3 Comparison with Previous

Approaches

When compared with previous methods in brain

tumor classiﬁcation, the proposed model provides

both high performance and excellent interpretabil-

ity through LRP visualizations. Traditional meth-

ods may lack interpretability or offer limited visual

insights into model predictions. The proposed ap-

proach, by contrast, offers a comprehensive under-

standing of how the model arrives at its predictions,

which is crucial in medical applications where trust

in the model’s decision-making is essential.

4.4 Interpretation of Results and

Implications

The results underscore the ability of the proposed

model to focus on clinically relevant regions, offer-

ing interpretability through heatmaps and overlayed

heatmaps. This helps clinicians to trust the model’s

predictions and provides assurance that the areas

identiﬁed by the model correspond to the tumor re-

gions in the MRI images. Additionally, the simpliﬁed

visualization without heatmaps ensures that clinicians

can rely on accurate, quantitative data when making

decisions. The report generated as shown in Figure 6

at the end of the process also facilitates patient care

by summarizing important tumor-related information

and offering treatment recommendations based on the

tumor type.

Figure 10: PDF Report of Meningioma Tumor

Figure 11: PDF Report of Pituitary Tumor

Figure 12: PDF Report of Glioma Tumor

4.5 Limitations of the Current

Approach

While the proposed model performs well in classify-

ing brain tumors and providing interpretability, there

are some limitations. One challenge arises when the

relevance is distributed across multiple areas, which

can reduce the clarity of heatmap visualizations. In

such cases, the model’s focus may become ambigu-

ous, making it harder to interpret the results clearly.

Additionally, although the heatmaps provide valuable

insights, they may not always offer the level of preci-

sion required in all medical applications, particularly

when multiple tumors or complex cases are involved.

INCOFT 2025 - International Conference on Futuristic Technology

Figure 13: PDF Report of No Tumor

5 CONCLUSION AND FUTURE

WORK

The proposed approach bridges the gap by applying

Layer-wise Relevance Propagation (LRP) to gener-

ate heatmaps that explain model predictions, enabling

medical practitioners to verify the rationale behind AI

outputs. This is crucial for understanding tumor prop-

erties, which vary in size, location, and type. LRP en-

hances the clarity and reliability of the system, help-

ing clinicians make better decisions, develop person-

alized treatment plans, and increase trust in AI-driven

diagnostic tools. Using ResNet-18 and LRP, the sys-

tem classiﬁes brain tumors in MRI scans into four cat-

egories: glioma, meningioma, pituitary, and no tumor.

The LRP results are intuitive and easy to understand,

making the system suitable for medical use. Fu-

ture work could improve performance by expanding

the dataset and exploring more complex architectures,

such as EfﬁcientNet (Litjens et al., 2017b). Incorpo-

rating multi-modal data and reﬁning LRP for clearer

visual explanations could further enhance model ac-

curacy. An intuitive interface with real-time predic-

tions could help healthcare providers quickly and ac-

curately identify patients in clinical settings.

REFERENCES

Babu Vimala, B., Srinivasan, S., Mathivanan, S., et al.

Detection and classiﬁcation of brain tumor using

hybrid deep learning models. sci rep 13, 23029

(2023).

Bach, S., Binder, A., Montavon, G., Klauschen, F.,

uller, K.-R., and Samek, W. (2015a). On pixel-

wise explanations for non-linear classiﬁer deci-

sions by layer-wise relevance propagation. PloS

one, 10(7):e0130140.

Bach, S., Binder, A., Montavon, G., Klauschen, F.,

uller, K.-R., and Samek, W. (2015b). On pixel-

wise explanations for non-linear classiﬁer deci-

sions by layer-wise relevance propagation. PloS

one, 10(7):e0130140.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep

residual learning for image recognition. In Pro-

ceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 770–778.

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J.,

and Maier-Hein, K. H. (2021). nnu-net: a

self-conﬁguring method for deep learning-based

biomedical image segmentation. Nature meth-

ods, 18(2):203–211.

Kulkarni, S. M. and Sundari, G. (2020). A frame-

work for brain tumor segmentation and classiﬁ-

cation using deep learning algorithm. Interna-

tional Journal of Advanced Computer Science

and Applications, 11(8).

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., Van Der Laak, J. A.,

Van Ginneken, B., and S

anchez, C. I. (2017a). A

survey on deep learning in medical image analy-

sis. Medical image analysis, 42:60–88.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., Van Der Laak, J. A.,

Van Ginneken, B., and S

anchez, C. I. (2017b). A

survey on deep learning in medical image analy-

sis. Medical image analysis, 42:60–88.

Pang, T., Li, P., and Zhao, L. (2023). A survey on

automatic generation of medical imaging reports

based on deep learning. BioMedical Engineering

OnLine, 22(1):48.

Ronneberger, O., Fischer, P., and Brox, T. (2015).

U-net: Convolutional networks for biomedi-

cal image segmentation. In Medical image

computing and computer-assisted intervention–

MICCAI 2015: 18th international conference,

Munich, Germany, October 5-9, 2015, proceed-

ings, part III 18, pages 234–241. Springer.

Vankdothu, R. and Hameed, M. A. (2022). Brain tu-

mor mri images identiﬁcation and classiﬁcation

based on the recurrent convolutional neural net-

work. Measurement: Sensors, 24:100412.

Layer-Wise Relevance Propagation for Classifying Brain MRI Images