This research project structures different enhanced architectures and models of CNNs using in particular the VGG16 model, for its featured simplicity and efficiency along with its pre-trained wights on ImageNet. The VGG16 models are well trained using transfer learning mechanism in fine-tuning the architecture on the ISIC2018 Task3 dataset. Then, the models are projected for skin cancer image classification in highlighting the state-of-the-art performance.
Deep learning models have showed great capabilities in data modelling on the various applications of image processing, including segmentation, classification, tagging, and many others. In particular, convolutional neural network (CNNs) has proved to be effective in capturing deep features on unstructured data that are well sited in the state-of-the-art. It is well competitive in comparison to the traditional algorithms of machine learning.
Table of Contents
1. Introduction
2. Related Work
3. Background
3.1 Deep Learning in Medical Diagnosis
3.2 Deep Learning architectures
3.3 Convolutional Neural Networks (CNNs)
3.3.1 Convolution Layer (C1, C3, and C5)
3.3.2 Pooling Layer (S2 and S4)
3.3.3 FCL (F6)
3.4 Transfer Learning (TL)
3.5 ISIC Dataset
3.5.1 Remarks of detecting a melanoma:
4. Approach
4.1 Pre-processing
4.1.1 Image Resizing
4.1.2 Data augmentation
4.1.2 Data split
4.2 Algorithm
4.2.1 VGG16 LL (Last layer fine-tuning)
4.2.2 VGG16 FL (First layer fine-tuning)
4.2.3 VGG16 ML (Middle layers fine-tuning)
4.2.4 VGG16 FU (Full layers fine-tuning)
4.3 Evaluation Metrics
1) Accuracy (A)
2) Log Loss (LL)
3) Confusion matrix (CM)
5. Experiments and Results
6. Conclusion
6.1 Future Works
6.2 Challenges
Research Objectives and Topics
The primary goal of this research project is to develop and evaluate enhanced convolutional neural network (CNN) architectures for the automated classification of medical images, specifically skin lesions. Using the VGG16 model as a baseline, the study explores the effectiveness of various transfer learning techniques, including different fine-tuning and layer-freezing strategies, to optimize performance on the ISIC2018 dataset.
- Automated skin lesion classification using deep learning.
- Comparative analysis of different CNN architectures (VGG16 variants).
- Implementation of transfer learning and fine-tuning strategies.
- Evaluation of preprocessing methods and data augmentation.
- Analysis of diagnostic performance using accuracy, log loss, and confusion matrices.
Excerpt from the Book
3.5.1 Remarks of detecting a melanoma:
There are two important methods to detect a melanoma. These methods are important to understand the necessary features that would help either manual or automated ML or DL in detecting and classifying the 7 types of melanoma lesions. Three-point checklist ABB is a typical method for diagnosis of melanoma and skin lesion, which is provided in the following [2]:
• Asymmetry: symmetry of color and structure in one or two perpendicular axes
• Atypical network: pigment network with irregular holes and thick lines
• Blue-white structures: any type of blue and/or white color, for example, combination of blue-white veil and regression structures.
Additionally, ABCD parameters method is also common for diagnosis of melanoma [2]:
• Asymmetrical shape: melanoma lesions are typically asymmetrical.
• Borders: melanoma lesions have the irregular border.
• Color: the presence of more than one color in melanoma lesions.
• Diameter: melanoma lesions are typically larger than 6mm in diameter.
Summary of Chapters
1. Introduction: This chapter introduces the role of deep learning in medical image processing and defines the specific aim of the internship: applying pre-trained VGG16 models to the ISIC2018 skin lesion dataset.
2. Related Work: This chapter reviews existing research articles that utilize ISIC challenges, detailing various deep learning architectures, preprocessing strategies, and the challenges faced by previous researchers in medical image classification.
3. Background: This chapter explains fundamental concepts including deep learning in medical diagnostics, the architecture of CNNs, the mechanics of transfer learning, and provides a descriptive overview of the ISIC2018 dataset.
4. Approach: This chapter details the methodology, including data preprocessing (resizing and augmentation), the configuration of the VGG16 model variations used, and the evaluation metrics (accuracy, log loss, confusion matrix) employed to assess the models.
5. Experiments and Results: This chapter presents the empirical outcomes of the four VGG16 architectures tested, comparing their performance metrics and highlighting the effectiveness of full-layer fine-tuning in achieving higher accuracy.
6. Conclusion: This final chapter summarizes the research findings, notes the persistent challenge of overfitting, and suggests future improvements such as incorporating more training data and adapting better training strategies.
Keywords
Convolutional Neural Network (CNN), Deep Learning (DL), Image Classification, Image Processing, Transfer Learning, VGG16, Medical Imaging, Skin Cancer, Melanoma, ISIC2018, Fine-tuning, Data Augmentation, Model Convergence, Overfitting, Diagnostic Accuracy.
Frequently Asked Questions
What is the core focus of this research project?
The project focuses on using deep learning, specifically convolutional neural networks (CNNs), to automate the classification of medical skin lesion images from the ISIC2018 dataset.
What are the central thematic areas?
The study centers on medical image diagnostics, the application of transfer learning in deep learning models, data preprocessing techniques, and comparative performance evaluation of different CNN configurations.
What is the primary goal of the study?
The goal is to improve the classification accuracy of skin lesion images by fine-tuning pre-trained VGG16 model architectures under four distinct experimental settings.
Which scientific methodology is employed?
The research uses the VGG16 architecture, applying various levels of fine-tuning (freezing/unfreezing different layers) and transfer learning, and evaluates the models using accuracy, log loss, and confusion matrices.
What is covered in the main section of the paper?
The paper covers the theoretical background of CNNs and transfer learning, a systematic approach to data handling (resizing, augmentation, splitting), and detailed experimental results comparing four VGG16 settings.
Which keywords characterize this work?
Key terms include CNN, Deep Learning, Image Classification, Transfer Learning, VGG16, Skin Cancer, ISIC2018, and Fine-tuning.
What specific role does the VGG16 model play in this research?
VGG16 serves as the baseline architecture. The researcher modifies it into four versions to test how freezing different blocks (layers) of the network affects the model's ability to learn from the specialized medical dataset.
How was the overfitting problem addressed during experiments?
The author noted significant overfitting after the 20th epoch across all models. While several standard techniques were used to minimize it, the author acknowledges that it remained a persistent challenge throughout the research period.
What were the main constraints encountered during the internship?
The primary constraints were limited computational resources (the author had to rely on a personal laptop rather than laboratory equipment) and difficulties in setting up the appropriate software environment for the deep learning algorithms.
What is the conclusion regarding the best-performing model?
The experiment concluded that the "VGG16 FU" (Full layers fine-tuning) model outperformed other variants, achieving a total accuracy of 90% on both the validation and test datasets.
- Citar trabajo
- Marwan Al Omari (Autor), 2021, Use of CNNs for the Classification of Medical Images, Múnich, GRIN Verlag, https://www.grin.com/document/1215084
