Major project 2
Project synopsis
Dermatological disease detection
using ResNet50 and XAI
Submitted by
Specialisation SAP ID Name
AIML 500092142 Aryan Mohan
AIML 500094143 Rishabh Sajwan
AIML 500094174 Abhishek Anand
AIML 500090985 Deepanshu Bisht
Project Mentor Cluster Head
Dr. Divya Mishra Dr. Anil Kumar
INDEX
S No. Content Page No.
1 Abstract 1
2 Introduction 2
3 Literature Review 2
4 Problem Statement 3
5 Objective 3
6 Methodology 4
7 PERT Chart 4
8 System Design 5
9 Model Used 5
10 Conclusion 5
11 References 5
1. Abstract
Skin diseases affect millions of people worldwide, with conditions ranging from common acne to
life-threatening melanoma and psoriasis. The challenge we face here, Detecting it in early stages
to improve treatment outcomes, yet access to dermatologists remains limited, especially in remote
and underserved areas..
This project leverages Deep Learning, specifically ResNet50, to develop a powerful AI-driven
system for automated skin disease detection. ResNet50, a Convolutional Neural Network (CNN), is
fine-tuned on extensive dermatological datasets to classify different skin conditions with high
accuracy. To make the predictions trustworthy and interpretable—we will take the use of
Explainable AI (XAI). By using Grad-CAM (Gradient-weighted Class Activation Mapping), the
model generates heatmaps that visually highlight the areas of an image most influential in its
decision-making.
The result is an application that allows users to upload skin lesion images and receive instant
AI-driven diagnostic insights. Designed with healthcare professionals and general users in mind,
the tool serves as a clinical decision support system, enhancing dermatological assessment and
facilitating early detection—particularly in resource-limited settings.
With the fusion of Deep Learning and Explainable AI, this project aims to revolutionize
dermatology by making advanced diagnostic tools more accessible, reliable, and interpretable.
Artificial Intelligence (AI) offers a powerful tool to address these challenges by enabling automated
diagnosis through deep learning models. These models can analyze images of skin lesions and
predict the presence of various skin diseases. This project aims to develop a practical and accessible
solution by integrating AI-based skin disease detection into an application that can be used by
healthcare professionals and individuals alike.
2. Introduction
Dermatology relies on visual assessment for diagnosing skin diseases, but access to dermatologists
is often limited, particularly in rural areas. Delays in diagnosis can lead to severe health
consequences, especially for conditions like melanoma that require early detection.
Recent advances in Artificial Intelligence (AI) and Deep Learning have enabled the development
of automated diagnostic tools that provide fast and accurate skin disease classification. This
project leverages ResNet50, a powerful Convolutional Neural Network (CNN), to classify different
skin conditions with high precision.
To enhance trust and transparency, the system integrates Explainable AI (XAI) techniques,
specifically Grad-CAM, which highlights image regions crucial for the model’s predictions. This
ensures interpretability, making AI-based diagnosis more reliable for medical professionals.
The AI model will be deployed as a desktop application, allowing users to upload skin lesion
images and receive instant diagnostic feedback. This solution aims to bridge the gap in
dermatological healthcare, ensuring early detection, improved accessibility, and better patient
outcomes.
Fig 1.1: Sequence Diagram of Project
3. Literature Review
● Deep learning has revolutionized medical image analysis, with Convolutional Neural
Networks (CNNs) demonstrating dermatologist-level accuracy in skin disease
classification.
Studies such as Esteva et al. (2017) showed that CNNs trained on large
dermatological datasets can achieve performance comparable to human experts.
Among the various
architectures, ResNet (He et al., 2016), VGGNet (Simonyan & Zisserman, 2015),
and InceptionNet (Szegedy et al., 2015) have been extensively studied for automated
skin disease classification.
● One of the most widely used benchmark datasets in dermatology research is HAM10000
(Tschandl et al., 2018), which contains over 10,000 dermatoscopic images covering
multiple types of pigmented lesions, including melanoma, basal cell carcinoma, and benign
nevi.
Research utilizing this dataset has demonstrated that pre-trained models using transfer
learning achieve high classification accuracy, with models like ResNet50 outperforming
earlier architectures due to its ability to extract deep hierarchical features while mitigating
vanishing gradient problems.
● Despite these advancements, interpretability remains a key challenge in clinical AI adoption.
Explainable AI (XAI) techniques, such as Grad-CAM (Selvaraju et al., 2017), have
been introduced to visualize regions contributing to model predictions, enhancing trust
and transparency in automated dermatological diagnosis. Such techniques are crucial for
regulatory approval and practical deployment in healthcare environments.
● However, several limitations persist. Dataset bias and class imbalance can lead to
overfitting and reduced generalizability in real-world applications. Studies suggest that data
augmentation (Perez & Wang, 2017), domain adaptation techniques, and ensemble
learning can improve robustness. Moreover, research indicates that incorporating
multi-modal data, such as patient history and clinical notes (Liu et al., 2020),
alongside image-based CNN models, can enhance diagnostic accuracy further.
● This project builds on these findings by implementing ResNet50 with transfer
learning, integrating Grad-CAM for explainability, and employing data
augmentation and cross-validation techniques to ensure generalizability and
reliability in dermatological disease detection.
● 4. Problem Statement
Healthcare systems in many parts of the world are under constant strain, especially in the
dermatology domain. Dermatological conditions are among the most common reasons for
medical consultations, yet there is a shortage of trained dermatologists. This problem is
particularly acute in rural and underserved areas, where patients must travel long distances
to access specialized care.
Manual diagnosis also presents challenges related to accuracy and consistency. Diagnoses
can vary among practitioners, and early signs of conditions like melanoma are often subtle
and easy to miss. These limitations highlight the need for automated diagnostic tools that
can provide fast and reliable assessments.
This project aims to bridge this gap by developing an AI-based system that can detect and
classify various skin diseases using images. By deploying this system as a desktop
application, we aim to make the solution accessible even in low-resource settings,
improving early detection and treatment outcomes.
5. Objectives
● Develop a deep learning-based model using ResNet50 to accurately classify
multiple dermatological conditions.
● Integrate Explainable AI (XAI) techniques, specifically Grad-CAM, to
improve interpretability and trust in model predictions.
● Design and deploy a user-friendly desktop application for easy accessibility
by healthcare professionals and individuals.
● Ensure robust performance and generalizability through data augmentation,
transfer learning, and rigorous validation.
● Provide real-time diagnostic support to enhance early disease detection and
treatment planning.
● Address dataset biases and class imbalance by implementing
preprocessing techniques and model fine-tuning.
● Optimize the efficiency and scalability of the system for practical deployment
in real-world clinical and non-clinical settings.
6. Methodology
+ Data Collection
The dataset used for this project includes the HAM10000 dataset and additional images
representing common skin conditions such as warts, psoriasis, and eczema. The data is divided into
training, validation, and testing sets to ensure comprehensive evaluation.
+ Data Preprocessing
Preprocessing steps include resizing images to 224x224 pixels, normalizing pixel values, and
applying data augmentation techniques such as rotation, zooming, and horizontal flipping. These
steps help improve the model's generalization to unseen data.
+ Model Development
We selected ResNet50 for this project due to its ability to learn deep representations while avoiding
the vanishing gradient problem. The model is pre-trained on ImageNet and fine-tuned on our
dataset. Additional dense layers are added to adapt the model for multi-class classification.
+ Training and Evaluation
The model is trained using the Adam optimizer with a learning rate of 0.0001. We use categorical
cross-entropy as the loss function and early stopping to prevent overfitting. Performance is
evaluated using accuracy, precision, recall, and F1-score.
7. PERT Chart/ Gantt Chart
Tas Task Name Description Predecess Duration
k ID o r(s) (Days)
1 Project Planning Define objectives, scope, and requirements - 3
2 Data Collection Acquire HAM10000 dataset and additional 1 5
images
3 Data Preprocessing Resize images, normalize data, handle class 2 4
imbalance
4 Model Selection Choose CNN architectures (ResNet, VGG, 3 2
InceptionNet)
5 Model Training & Train model with augmented data, evaluate 4 7
Evaluation accuracy
6 Explainability Implement Grad-CAM for interpretability 5 3
Integration (XAI)
7 Deployment Convert model to ONNX/TensorFlow Lite for 6 3
Preparation UI use
8 Frontend Development Build a UI for image input and predictions 7 5
9 Backend/API Set up API for model inference 8 4
Development
10 Testing & Optimization Validate system performance, optimize speed 9 6
& memory
11 Final Deployment Deploy system on cloud/local machine 10 3
12 Documentation & Create final documentation and research 11 4
Report paper
8. System Design
The system consists of several components:
1. User Interface: A Tkinter-based desktop application for image input and prediction display.
2. Backend Model: The trained ResNet50 model loaded using TensorFlow.
3. Prediction Module: Processes the input image and returns the predicted class and confidence
score.
4. Explainable AI: Visualizes heatmaps to show important regions contributing to the prediction.
Model Used
Fig 1.2 : Model UML Diagram
The ResNet50 model serves as the backbone of this project. It is a deep CNN architecture with
residual connections that allow the model to learn complex features efficiently. Transfer learning is
used to fine-tune the model on our dataset, achieving high accuracy even with limited data.
9. Conclusion
This project demonstrates the potential of AI in transforming dermatology by enabling automated
skin disease detection. The use of CNNs provides a reliable tool for early diagnosis, while XAI
ensures transparency and interpretability. With further testing and real-world deployment, thi
system can improve healthcare accessibility and outcomes for patients worldwide.
10. References
1. Tschandl, P., et al. (2018). "The HAM10000 Dataset: A Large Collection of Multi-Source
Dermatoscopic Images of Common Pigmented Skin Lesions."
2. He, K., et al. (2016). "Deep Residual Learning for Image Recognition."
3. Selvaraju, R.R., et al. (2017). "Grad-CAM: Visual Explanations from Deep Networks via
Gradient-based Localization."
4. Litjens, G., et al. (2017). "A Survey on Deep Learning in Medical Image Analysis."
