Explainable AI Framework for Precise and Trustworthy Skin Cancer

Diagnosis

Vandana Kate

, Arohi Kate, Chanchal Bansal, Charu Pancholi and Ashvini Patidar

Department of CSIT, Acropolis Institute of Technology and Research, MP, India

Keywords:

Skin Cancer, Explainable AI (XAI), Deep Learning, Inception-ResNet V2, Grad-CAM, Clinical Decision

Support, Skin Lesion Classiﬁcation, Non-Invasive Diagnosis.

Abstract:

Skin cancer, most especially melanoma, is a recognized health issue across the world and its management

depends on early and correct diagnosis.Conventional methods like biopsies are relatively precise and reliable

but they are time consuming and invasive and may cause either an infection or an outbreak. Non-invasive

procedures such as dermoscopy depend on the knowledge of the physician, which can cause variability and

randomness. To address these challenges, we propose an explainable AI (XAI) framework for precise and

trustworthy skin cancer diagnosis. Our model integrates VGG16, InceptionV3, Inception-ResNet V2 and

DenseNet-201 deep learning architectures ﬁne-tuned on the HAM10000 benchmark dataset to distinguish skin

lesions as benign or malignant. To ensure transparency and trust in the model’s predictions, we incorporate

cutting-edge explainability techniques, including LIME (Local Interpretable Model-agnostic Explanations),

SHAP (SHapley Additive exPlanations) and gradient-based methods like Grad-CAM. These tools highlight

key image features and regions that inﬂuence model decisions. This proposed work deepens the knowledge in

the ﬁeld of using AI in the diagnosis of skin cancer and paves the way for integrating explainability into AI

healthcare systems, improving accuracy and user trust.

1 INTRODUCTION

Skin Carcinoma is one of the most commonly diag-

nosed and eventually fatal types of cancer that con-

tinuously on the rise all over the world. Although

deep learning models for more complex layers are, in

fact, highly accurate for diagnosing skin cancer, clin-

icians often struggle to understand how models draw

their conclusions, which is a major drawback of popu-

lar AI systems and existing approaches to skin cancer

diagnosis (P. Linardatos and Kotsiantis, 2020),(Wang

et al., 2021). To address this challenge, Explainable

AI (XAI) approaches have been developed, which can

be categorized as follows:

1.1 Function Based Approach

This approach focuses on understanding the inner

workings of the model, such as its number of lay-

ers, parameters, and hyperparameters. Both saliency

maps and feature importance are simple procedures

https://orcid.org/0000-0002-2281-2187

that explain how features affect the output of a model.

An example is Grad-CAM (Gradient-weighted Class

Activation Mapping), which helps in visualizing

which of the areas in an image contributes to a

model’s decision.

1.2 Result Based Approach

This approach gives explanations for speciﬁc predic-

tions based on approximations of the complex model

with models that are easier to understand. Techniques

like LIME (Local Interpretable Model-agnostic Ex-

planations) and SHAP (SHapley Additive exPlana-

tions) are popular for breaking down individual pre-

dictions, offering insights into which features con-

tributed to the outcome. For instance, LIME can be

used to explain predictions in text classiﬁcation tasks

by generating interpretable models around a speciﬁc

instance (K. Aas and Løland, 2021).

260

Kate, V., Kate, A., Bansal, C., Pancholi, C. and Patidar, A.

Explainable AI Framework for Precise and Trustworthy Skin Cancer Diagnosis.

DOI: 10.5220/0013590500004664

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

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 260-267

ISBN: 978-989-758-763-4

1.3 Conceptual Based Approach

This approach looks forward to providing human in-

terpretable representation based on high-level fea-

tures or domain knowledge. For instance, the con-

ceptual explanation in medical imaging can be best

illustrated by the ability to depict images that contain

”tumor-like” structures or ”abnormal patterns”. This

type of explanation is the most helpful in ﬁelds where

domain knowledge is highly important.

1.4 Mixed Approach

The mixed approach combines elements from the

above approaches to provide a comprehensive expla-

nation. Using function-based, result-based, and con-

ceptual explanations, this approach provides a better

understanding of the choice of model decisions. For

example, in complex tasks such as medical diagnosis,

a mixed approach might combine Grad-CAM visual-

izations with LIME-based feature importance scores,

making the behavior of the model more comprehensi-

ble and understandable to the human audience (Lund-

berg, 2017). Such approaches assist in explaining and

verifying mainly used AI solutions in such critical ar-

eas such as medicine and self autonomous systems.

The proposed work aims to expand the horizons of

health care, utilizing current standard methodologies

such as SHAP, LIME, and Grad-CAM while perform-

ing layer-wise analysis to assess and interpret CNN

model decisions for skin cancer diagnosis. The goal

is to make the decision process clearer to understand

and easier to control since it provides quantitative and

qualitative data of how each pixel or region impacts

the decision making of the developed model.

1.5 Objectives

The schematic diagram of the proposed approach is

shown in ﬁgure 1 with various objectives as follows.

i. Perform an exploration of diverse interpretabil-

ity techniques and compare and contrast them ac-

cording to their suitability to explain CNN deci-

sions.

ii. Incorporate layer-wise analysis in the propagation

of information through the network in an effort to

isolate the level of participation of every layer in

the ﬁnal result of the model.

iii. Promote the interpretation of mammography di-

agnostic results in order to improve the level of

trust of physicians and patients.

Figure 1: Schematic Diagram of the Proposed Model

2 RELATED WORKS OF XAI ON

DIFFERENT APPLICATION

DOMAINS

Partial Dependence (PD) plots are criticized for

failing when it comes to interpreting black-box mod-

els with correlated predictors, which is why (Apley

and Zhu, 2020) presented the Accumulated Local Ef-

fects (ALE) plots as more robust and less time- con-

suming. ALE plots do not exhibit extrapolation prob-

lems and are suggested as a standard tool for inter-

pretability in supervised learning.

Machine learning models including SVM, DNN, and

Random Forest were used by (P. Jain and Jain, 2024)

to predict PCOS with 97% accuracy and used the two

XAI strategies, namely LIME and SHAP to analyze

key determinants of the disorder.

A recent study conducted by (K. Wei and Chen,

2022) implemented XAI in agricultural classiﬁca-

tion through the deep learning models of VGG,

GoogLeNet, ResNet, in the fruit leaves dataset.

During training the ResNet model, an accuracy of

(99.11%, 99.4%, 99.89%) was achieved, and the at-

tention module improved feature extraction from the

input images and provided information about what as-

Explainable AI Framework for Precise and Trustworthy Skin Cancer Diagnosis

261

pect of the image the model is focusing on while clas-

sifying.

(H. Naeem and Ullah, 2022) proposed an AI-based

explainable approach for malware detection using IoT

devices using a ﬁne-tuned Inception-v3 CNN model

with transfer learning. By using color image malware

display of Android Dalvik Executable File (DEX),

the model achieved 98.5% accuracy in binary classi-

ﬁcation and 91% in multiclass prediction, surpassing

other methods in various evaluation metrics.

(Molle et al., 2018) represented the dermatology case,

where they observed that CNNs inspect features that

are similar to those examined by dermatologists for

skin lesions; however, more analysis is required for

the interpretation of convolutional neural networks.

(J. M. Rozanec and Mladenic, 2022) proposed a

Knowledge Graph-based XAI architecture that is used

for demand forecasting with conﬁdential high-level

explanations and actions based on domain knowledge

while preserving sensitive model details.

Four attribution methods were evaluated by (F. Ei-

tel and the Alzheimer’s Disease Neuroimaging Initia-

tive (ADNI), 2019) for CNN-based Alzheimer’s clas-

siﬁcation based on MRI data. It also clearly indicated

that there are large ﬂuctuations, while guided back-

propagation and LRP yielded the most consistent val-

ues; so, it is necessary to use domain-speciﬁc criteria

instead of a visual assessment of the maps.

(S. Pereira and Silva, 2018) introduced the idea of em-

ploying CNNs to detect the grade of glioma solely

based on MRI data, thus avoiding the need for a

biopsy. They assessed prognosis using whole brain

and automatic tumor areas and used interpretability

methods to guide the models to concentrate on the re-

gions that are indicative of tumor grade.

(Mehta and Passi, 2022) used XAI for hate speech

detection including pre-processing and exploratory

analysis of datasets. LSTM achieved an accu-

racy of 97.6% on the Google Jigsaw dataset, while

BERT variants (BERT + ANN: 93.55%, BERT +

MLP: 93.67%) were evaluated for explainability us-

ing LIME and the ERASER (Evaluating Rationales

and Saliency for Explanations in Reasoning) meth-

ods.

(S. Y. Lim and Lee, 2022) extended the XAI tech-

niques of image classiﬁcation to deepfake audio de-

tection, providing an understanding of interpretability

and explanation of model decisions involving varia-

tions of pitch and rhythm. The ﬁndings emphasized

that the interpretability was consistent across environ-

ments and noted its divergence between human and

model perceptions provided information to respond to

the emerging problem of generative fake media.

(Kim and Joe, 2022) proposed an XAI approach for

deep learning self-driving car models that maps image

regions that have signiﬁcant impacts on CNN deci-

sion making using sensitivity analysis. This increases

reliability in conjunction with the application of the

devices.

LSTM, Bi-LSTM, and Bi-GRU-LSTM-CNN mod-

els were employed by (A. Adak and Alamri, 2022)

for sentiment analysis of FDS reviews with accuracy

rates equal to 96.07%, 95.85%, and 96.33%, corre-

spondingly. LSTM was chosen for false negatives as

they are lower compared to the other. The two XAI

methods that we used were SHAP and LIME; which

provided explanations by isolating the words most in-

ﬂuential to the sentiment of the models.

3 DATASET- SKIN CANCER

MNIST (HAM10000)

The HAM10000 dataset (HAM, ) contains more

than 10,000 dermatoscopic images of skin lesions,

mainly melamonous to diagnose and categorize skin

cancer. The dataset has seven different classes,

namely: Melanocytic nevi, Melanoma, Benign ker-

atosis similar lesions, Basal cell carcinoma, Actinic

keratoses, Vascular lesions, and Dermatoﬁbroma.

This data set is relatively difﬁcult in the development

of models due to its applicability, especially due to

the high imbalance of classes. In this regard, it plays

a vital reference for a more accurate diagnosis of less

common but potentially serious skin diseases.

4 DEEP CONVOLUTIONAL

NEURAL NETWORKS (CNN)

FOR IMAGE CLASSIFICATION

Deep CNNs are a well established deep learning

architecture most applicable to image classiﬁcation

problems. Due to their capacity to learn about hier-

archical features of items, Convolution Neural Net-

works are particularly useful in the skin cancer image

classiﬁcation task. In the initial layers model works

with simple features or basic or low-level features

such as edges or shapes, and as one passes through

the network, the high-level or more abstract features

are extracted in the latter layers and then the network

is able to differentiate between the different types of

skin cancer lesions.

INCOFT 2025 - International Conference on Futuristic Technology

262

4.1 Experimental Setup: Baseline and

Pre-trained Models

The work proposed here incorporates both the base-

line CNN and ﬁne-tuned pre-trained deep models as

described below:

1. Baseline CNN: This model acts as a benchmark

for comparison. It uses only three convolution and

pooling layers and has a basic structure in order to

create a base for future models.

2. Fine-tuned Pretrained Models: Four pre-trained

models – VGG16, Inception V3, Inception

ResNet V2 and DenseNet 201, were applied to

the HAM10000 skin cancer data set. These

models were initially trained on a huge dataset

like ImageNet, allowing them to extract/learn

generic image features. The higher layers of

these modnd.els were ﬁne-tuned to the special-

ized task of skin lesion segmentation based on the

HAM10000 dataset. This approach takes advan-

tage of the existing knowledge of the pre-trained

models but modiﬁes them according to the classi-

ﬁcation problem at hand. Various models used are

brieﬂy described below-

• VGG16: A well-known pre-trained deep

model with a hierarchical convolutional pool-

ing architecture. It is composed of ﬁve con-

volutional blocks, where third, fourth, and ﬁfth

blocks have four convolutional layers. This ar-

chitecture enables the model to capture increas-

ingly complicated features, making it highly ef-

fective for image classiﬁcation tasks.

• Inception V3: This top ImageNet model uti-

lizes ”Inception modules” to efﬁciently extract

features. Speciﬁcally, it is a computationally

efﬁcient and effective framework for feature ex-

traction.

• Inception ResNet V2: The model is based on

Inception V3 and combines residual connec-

tions to enable deeper networks training. In-

ception ResNet V2 also minimizes some possi-

ble issues of Batch Normalization.

• DenseNet 201: This model comprises four

”dense blocks”, where each block generates

feature maps through a series of operations

(batch normalization, ReLU, and a 3x3 con-

volution). The layers between dense block are

called transition layers which consist of convo-

lution and pooling layers.

The ﬁgure 2 illustrates the training and validation ac-

curacy and loss graphs for the above four models, pro-

viding a tabular summarization of various model per-

formance during training.

Figure 2: Model Training Performance

5 EXPLAINABLE MACHINE

LEARNING

Healthcare AI is still struggling mainly because

it is hard to integrate something you cannot fully

explain with AI-made decisions that require pa-

tient trust, model interpretability, and feedback for

accurate and reliable results. In response to these

challenges, most of the proposed works make use

of CAM (Class Activation Mapping) methods that

identify the areas in an image that a model uses to

classify a given class, hence improving interactivity.

Some CAM techniques (Zhou et al., 2015) also

employ Global Average Pooling (GAP) or Global

Max Pooling (GMP) to preserve spatial structures and

detect discriminative areas, which improves decision-

making and reliability in the healthcare domain.

Although GMP analyzes the most signiﬁcant section

of an object, GAP can learn about and pinpoint the

presence of a complete object. GAP helps better

understand by including all discriminative factors

in consideration. GAP considers all discriminative

parts, ensuring a more comprehensive understanding.

Therefore, GAP has higher accuracy than GMP

in localization problems and is recommended for

precise spatial localization in AI-based healthcare

decision support systems.

Subsequent sections of this paper demonstrate the

experiments performed with the help of explanatory

methods such as LIME, SHAP, and Grad-CAM to

analyze the outcomes provided by deep learning

regarding the classiﬁcation of skin cancer. These

methods help to explain the model to justify the

Explainable AI Framework for Precise and Trustworthy Skin Cancer Diagnosis

263

kind of decision it made by pointing out features or

regions in the input images that contribute mostly

to the prediction and thus adds to reliability of the

model.

6 LIME (LOCAL

INTERPRETABLE

MODEL-AGNOSTIC

EXPLANATIONS) FOR

EXPLAINING SKIN CANCER

IMAGE INTERPRETATION

As experimented with skin cancer images, the

LIME algorithm 1 offers a way to recognize which

areas of an image are important for model decision. It

does so by applying some transformations/perturba-

tions to the input image (for example, deleting some

superpixels) and comparing the changes to the out-

put. LIME then ﬁts a simple linear model – or Deci-

sion Tree – to mimic the behavior of the complicated

deep model in the surroundings of the original image.

Since this linear model involves learning coefﬁcients,

the super-pixels that have big contributions towards a

prediction can be pinpointed, and how the model is

making decisions can be revealed by overlaying the

heat-map on the original image. Heatmap is a graphi-

cal data analysis tool in which data values are encoded

as colors. It is popular for displaying the density of

values on a matrix or a grid, where each cell will rep-

resent the value of the data point it contains.

Algorithm 1 LIME Algorithm for Skin Cancer Image

Predictions

Complex model M, skin lesion image I, predicted output

class y = argmax M(I). Feature importance and heatmap

visualization.

Step 1: Segment Image

Divide I into k superpixels {S

, S

, . . . , S

Step 2: Generate Perturbed Images

Create perturbed images by masking superpixels.

Step 3: Get Predictions

Pass each perturbed image through M to obtain predictions.

Step 4: Compute Weights

Compute similarity between I and perturbed images. As-

sign weights based on the similarity.

Step 5: Train Model

Fit a weighted linear regression model on the perturbed im-

ages.

Step 6: Create Heatmap

Assign model coefﬁcients to superpixels. Visualize the

heatmap highlighting regions inﬂuencing the prediction.

Figure 3: Visualizing Heatmaps for LIME

The image in ﬁgure 3 illustrates how different

segments (superpixels) of an image contribute to the

whole through the use of color. A positive coefﬁcient

value means that it has a direct proportional relation-

ship with the prediction of the model, and regions that

include brighter colors (like yellow) have a high pos-

itive coefﬁcient. However, the negative coefﬁcients

are associated with the darker color, particularly the

purple color to represent areas where it has a reduc-

ing impact on the conﬁdence of the model. The right

side of the ﬁgure presents a color bar that might help

us to understand the values of the coefﬁcients. Based

on the intensity of the color depicted on the heatmap,

the areas are delineated, which are most important for

decision making in the chosen model with reference

to the image.

7 SHAP ALGORITHM FOR SKIN

CANCER IMAGE

INTERPRETATION

SHAP (SHapley Additive exPlanations) values act

as uniﬁed/global measure in that they provide a mea-

sure of how each feature contributes progressively to

the model prediction across the given dataset. This

theory comes from game theory, where there is an

analysis of how each individual contributes differ-

ently to the team’s outcome. Likewise, SHAP values

explain how much each feature plays in the predic-

tion of the model, giving an equal chance for each

feature to contribute positively or negatively to the

model across all permutations of features in a dataset

as illustrated in algorithm 2. The color map on the

right-hand side of ﬁgure 4 indicates the SHAP val-

ues. The blue regions correspond to negative SHAP

values, while the red regions indicate positive SHAP

values. The SHAP values are overlaid on the origi-

nal image, highlighting the most important regions to

make the prediction.

Label: 1.0 corresponds to a malignant lesion with

a probability of 56% malignancy. The SHAP visu-

alization data reveal the fact that the model concen-

INCOFT 2025 - International Conference on Futuristic Technology

264

Algorithm 2 SHAP Algorithm for Skin Cancer Image

Predictions

Complex model M, skin lesion image I, background dataset

D. SHAP values for each pixel of I and visualization.

Step 1: Background Data

Choose representative background images D =

, I

, . . . , I

Step 2: Perturb Features

Divide I into k superpixels {S

, S

, . . . , S

Apply random binary masks to perturb the image.

Step 3: Compute Predictions

Generate samples by masking the background images and

passing them through M.

Step 4: Compute SHAP Values

Calculate SHAP values for each superpixel. The formula

calculates the average contribution of some feature j by

considering permutations of other features and measuring

how adding j changes the model’s prediction

Step 5: Aggregate SHAP Values

Assign SHAP values to corresponding pixels in each super-

pixel.

Step 6: Visualize SHAP

Create a heatmap using SHAP values, with positive contri-

butions in red and negative in blue.

trated on some peripheral regions illustrated by the

red points to predict malignancy.

Label: 0.0 corresponds to a benign lesion with

49% probability of malignancy. The SHAP visualiza-

tion shows that the model considered multiple areas

marked in blue (negative) and red (positive) across the

lesion when making the prediction.

Figure 4: Visualizing Heatmaps for SHAP

8 GRAD-CAM ALGORITHM FOR

SKIN CANCER IMAGE

INTERPRETATION

Grad-CAM supports deep learning solutions in

terms of interpretability, especially, in image classi-

ﬁcation problems that show tumor malignancy, as il-

lustrated in the algorithm 3. It allows practitioners,

such as radiologists or researchers, to:

• Ensure that the model targets areas that are of

Figure 5: Visualizing Heatmaps for GRAD CAM

medical importance.

• Find any shortcut such that the model might pay a

lot of attention to some part of the image which is

of no importance.

• Understand the decision-making process, espe-

cially suitable in industries such as healthcare,

where interpretability is essential.

The ﬁgure 5 provides a step by step explanation of

Grad-CAM applied to a skin cancer image as input,

with overlaid heatmaps.

Algorithm 3 Grad-CAM for Skin Cancer Image Pre-

dictions

Trained model M, image I, target class c, layer L.

Grad-CAM heatmap.

Step 1: Prediction

Pass I through M to get the prediction y

= M(I)[c].

Step 2: Gradients

Compute

∂y

∂A

for layer L.

Step 3: Weight Feature Maps

Calculate α

∑

∂y

∂A

i j

Step 4: Generate Heatmap

Compute H

i j

= ReLU



∑

i j



Step 5: Rescale Heatmap

Upsample H

i j

using bilinear interpolation to match

the size of the input image I . Scale H

i j

to the range

[0, 1] for visualization.

Step 6: Superimpose Heatmap Overlay H onto I to

visualize important regions.

9 RESULT

The graph in ﬁgure 6 represents the performance

of the model on the Interpretation over Union (IoU)

Explainable AI Framework for Precise and Trustworthy Skin Cancer Diagnosis

265

scale, which measures the overlap of the regions of

importance highlighted by explainability methods and

the actual areas of interest. X-axis shows the evalu-

ated models and the Y-axis shows the IoU which re-

ﬂects the accuracy to which the regions deﬁned by the

heatmaps correspond to the regions of interest. Val-

ues range from 0 to 1, where: 0: there is no over-

lap between predicted and true regions. 1: Perfect

overlap. Grad-CAM, when compared to both LIME

Figure 6: Comparison of IoU Scores for Explainable Meth-

ods: Grad-CAM, LIME, and SHAP

and SHAP, provides better results most of the time

because of its capability to visualize model decisions.

Some reasons for obtaining better results are:

1. Grad-CAM takes advantage of the overall

channel-wise architecture of the CNN, which is

designed to uncover spatial patterns at multiple

levels of abstraction. Grad-CAM focuses partic-

ularly on the last layers of the convolutional neu-

ral network and explains what components of the

input image are important for the model’s predic-

tion.

2. Grad-CAM emphasizes discriminative areas, such

as lesions or patterns that are medicinally impor-

tant.

3. Grad-CAM output consists of visual heatmaps

that are easily understandable, as they indicate

which areas inﬂuenced the model’s decisions.

These heatmaps provide excellent localization in

certain parts of the image.

4. Unlike other models, Grad-CAM uses gradients

of global feature maps, making it resistant to noise

within the input image and resulting in accurate

explanations.

5. LIME suffers from the limitation that it is effec-

tive at the local level but could be problematic

when addressing the global context in image data.

6. SHAP is limited in that it can elicit the impor-

tance of features but does not provide spatial in-

terpretations subsumed in the importance scores.

Grad-CAM, however, demonstrates high accuracy

in detecting spatial relevance.

10 SUMMARY OF USE CASES

By combining Grad-CAM, LIME, and SHAP, we

achieve a comprehensive explainability framework:

• Grad-CAMare useful for achieving a more visual

conﬁrmation of the areas of .focus

• LIMEcould be used to explain generic decision-

making behaviors.

• SHAP could be used to provide high-level pixel-

level information and summarize important char-

acteristics.

Such a combined strategy allows for the necessary

balance between high-level interpretability and de-

tailed analysis, which may be crucial in such ﬁelds as

medical diagnostics, for example, skin cancer recog-

nition.

11 CONCLUSION

The comparison of Grad-CAM with LIME and

SHAP across models such as CNN, VGG16, Incep-

tion V3, Inception ResNet V2, and DenseNet 201

shows that Grad-CAM has better results in terms of

IoU, demonstrating its ability to detect medically im-

portant zones. LIME and SHAP do not focus on

regions similar to ground truths, while Grad-CAM

concentrates on the class-discriminative areas and are

useful for healthcare-related applications with high

visual interpretability performance. Emerging archi-

tectures such as Vision Transformers (ViTs) and its

evaluation on a more heterogeneous sample of med-

ical imaging data could validate these ﬁndings. Ap-

plying domain knowledge and feedback from users

also holds the prospect for enhancing usefulness of

explainability methods in more important and signiﬁ-

cant applications.

REFERENCES

Ham10000 dataset, kaggle. [Online]. Available:

https://www.kaggle.com/datasets/kmader/skin-

cancer-mnist-ham10000.

A. Adak, B. Pradhan, N. S. and Alamri, A. (2022). Un-

boxing deep learning model of food delivery service

reviews using explainable artiﬁcial intelligence (xai)

technique. Foods, 11(14):2019.

Apley, D. W. and Zhu, J. (2020). Visualizing the effects

of predictor variables in black box supervised learning

models. Journal of the Royal Statistical Society Series

B: Statistical Methodology, 82(4):1059–1086.

INCOFT 2025 - International Conference on Futuristic Technology

266

F. Eitel, K. R. and the Alzheimer’s Disease Neuroimaging

Initiative (ADNI) (2019). Testing the robustness of

attribution methods for convolutional neural networks

in mri-based alzheimer’s disease classiﬁcation. In In-

terpretability of Machine Intelligence in Medical Im-

age Computing and Multimodal Learning for Clinical

Decision Support, page 3–11. Springer.

H. Naeem, B. M. A. and Ullah, F. (2022). Explainable arti-

ﬁcial intelligence-based iot device malware detection

mechanism using image visualization and ﬁne-tuned

cnn-based transfer learning model. Computational In-

telligence and Neuroscience, 2022:7671967.

J. M. Rozanec, B. F. and Mladenic, D. (2022). Knowl-

edge graph-based rich and conﬁdentiality preserving

explainable artiﬁcial intelligence (xai). volume 81,

page 91–102.

K. Aas, M. J. and Løland, A. (2021). Explaining individual

predictions when features are dependent: More accu-

rate approximations to shapley values. Artiﬁcial Intel-

ligence, 298:103502.

K. Wei, B. Chen, J. Z. S. F. K. W. G. L. and Chen, D. (2022).

Explainable deep learning study for leaf disease clas-

siﬁcation. Agronomy, 12(5):1035.

Kim, H. S. and Joe, I. (2022). An xai method for convolu-

tional neural networks in self-driving cars. PLOS One,

17(8):e0267282.

Lundberg, S. (2017). A uniﬁed approach to interpreting

model predictions. arXiv preprint, arXiv:1705.07874.

Mehta, H. and Passi, K. (2022). Social media hate

speech detection using explainable artiﬁcial intelli-

gence (xai). Algorithms, 15(8):291.

Molle, P. V. et al. (2018). Visualizing convolutional neural

networks to improve decision support for skin lesion

classiﬁcation. In Understanding and Interpreting Ma-

chine Learning in Medical Image Computing Applica-

tions, pages 115–123. Springer.

P. Jain, R. K. Mishra, A. D. and Jain, N. K. (2024). Xplain-

able ai for deep learning model on pcod analysis. In

XAI Based Intelligent Systems for Society 5.0, page

131–152. Elsevier.

P. Linardatos, V. P. and Kotsiantis, S. (2020). Explain-

able ai: A review of machine learning interpretability

methods. Entropy, 23(1):18.

S. Pereira, R. Meier, V. A. M. R. and Silva, C. (2018). Adap-

tive explainable ai for brain tumor segmentation. In

Medical Image Computing and Computer-Assisted In-

tervention – MICCAI 2018, page 92–100. Springer.

S. Y. Lim, D. K. C. and Lee, S. C. (2022). Detecting deep-

fake voice using explainable deep learning techniques.

Applied Sciences, 12(8):3926.

Wang, X. et al. (2021). Explainable deep learning for efﬁ-

cient and robust pattern recognition: A survey of re-

cent developments. Pattern Recognition, 120:108102.

Explainable AI Framework for Precise and Trustworthy Skin Cancer Diagnosis

267