Enhanced Disease Classification with HybridCNNViT model and Synthetic Oversampling (SMOTE)
(2025) STAN40 20251Department of Statistics
- Abstract
- Early and accurate classification of skin diseases plays a crucial role in improving diagnostic efficiency and patient outcomes. This thesis introduces HybridCNNViT, a deep learning algorithm that combines a setup of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to capture both local texture and global structure from skin lesion images. To address the common issue of class imbalance, where rare but critical skin conditions are underrepresented, the model applies the Synthetic Minority Oversampling Technique (SMOTE) during data preprocessing. Further enhancements include: attention mechanisms to help the model focus on relevant regions and regularization techniques to prevent overfitting. The training process also... (More)
- Early and accurate classification of skin diseases plays a crucial role in improving diagnostic efficiency and patient outcomes. This thesis introduces HybridCNNViT, a deep learning algorithm that combines a setup of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to capture both local texture and global structure from skin lesion images. To address the common issue of class imbalance, where rare but critical skin conditions are underrepresented, the model applies the Synthetic Minority Oversampling Technique (SMOTE) during data preprocessing. Further enhancements include: attention mechanisms to help the model focus on relevant regions and regularization techniques to prevent overfitting. The training process also benefits from dynamic learning rate scheduling and data augmentation strategies. Evaluated on a challenging dermatology dataset, the model achieves a test precision of 88.42% and a weighted F1 score of 0.8802, demonstrating competitive performance compared to existing methods and a strong potential for clinical application. (Less)
- Popular Abstract
- Early and accurate classification of skin diseases plays a crucial role in improving diagnostic efficiency and patient outcomes. This thesis introduces HybridCNNViT, a deep learning algorithm that combines a setup of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to capture both local texture and global structure from skin lesion images. To address the common issue of class imbalance, where rare but critical skin conditions are underrepresented, the model applies the Synthetic Minority Oversampling Technique (SMOTE) during data preprocessing. Further enhancements include: attention mechanisms to help the model focus on relevant regions and regularization techniques to prevent overfitting. The training process also... (More)
- Early and accurate classification of skin diseases plays a crucial role in improving diagnostic efficiency and patient outcomes. This thesis introduces HybridCNNViT, a deep learning algorithm that combines a setup of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to capture both local texture and global structure from skin lesion images. To address the common issue of class imbalance, where rare but critical skin conditions are underrepresented, the model applies the Synthetic Minority Oversampling Technique (SMOTE) during data preprocessing. Further enhancements include: attention mechanisms to help the model focus on relevant regions and regularization techniques to prevent overfitting. The training process also benefits from dynamic learning rate scheduling and data augmentation strategies. Evaluated on a challenging dermatology dataset, the model achieves a test precision of 88.42% and a weighted F1 score of 0.8802, demonstrating competitive performance compared to existing methods and a strong potential for clinical application. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9204692
- author
- Boehm, Kristina LU and Singh, Raghwendra LU
- supervisor
-
- Michal Kos LU
- organization
- course
- STAN40 20251
- year
- 2025
- type
- H1 - Master's Degree (One Year)
- subject
- keywords
- ViT (Vision Transformer), CNN (Convolutional Neural Network), SMOTE (Synthetic Minority Over-sampling Technique), CBAM (Convolutional Block Attention Module)
- language
- English
- id
- 9204692
- date added to LUP
- 2025-07-09 11:43:46
- date last changed
- 2025-07-09 11:43:46
@misc{9204692,
abstract = {{Early and accurate classification of skin diseases plays a crucial role in improving diagnostic efficiency and patient outcomes. This thesis introduces HybridCNNViT, a deep learning algorithm that combines a setup of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to capture both local texture and global structure from skin lesion images. To address the common issue of class imbalance, where rare but critical skin conditions are underrepresented, the model applies the Synthetic Minority Oversampling Technique (SMOTE) during data preprocessing. Further enhancements include: attention mechanisms to help the model focus on relevant regions and regularization techniques to prevent overfitting. The training process also benefits from dynamic learning rate scheduling and data augmentation strategies. Evaluated on a challenging dermatology dataset, the model achieves a test precision of 88.42% and a weighted F1 score of 0.8802, demonstrating competitive performance compared to existing methods and a strong potential for clinical application.}},
author = {{Boehm, Kristina and Singh, Raghwendra}},
language = {{eng}},
note = {{Student Paper}},
title = {{Enhanced Disease Classification with HybridCNNViT model and Synthetic Oversampling (SMOTE)}},
year = {{2025}},
}