(It is suggested to go through this link to have a detailed explaination of our project: https://github.com/DeRoyace/lightweight-skin-disease-classifier/blob/main/work/paper/Lightweight%20Deep%20Learning%20Approach%20for%20Skin%20Disease%20Classification.pdf)

---

Skin Disease Detection using Machine Learning

old Notebook Cancer model pox model

---

🎯 Target for Semester-2:

1. Using MobileNetV2
2. New dataset: HAM 10000, MPOX Lession
3. Increase skin disease classes - pox and cancer (total 13 classes).
4. Build an mobile app using flutter - to scan and detect skin disease instantly.

Explanation of Work

The research paper titled "Skin Disease Detection using Machine Learning" by Debkrishna Hatua, Rohit Roy and Sahil Ghosh presents a comprehensive methodology for detecting and classifying skin diseases through advanced image processing techniques and deep learning algorithms. The study emphasizes the importance of accurate diagnosis in dermatology, aiming to enhance the efficiency and effectiveness of skin disease identification.

Methodological Steps

1. Image Preprocessing:

   • Noise Removal: Applies Gaussian filtering to reduce noise and blurring in the images, ensuring clearer input for further processing.

2. Segmentation:

   • Employs the GrabCut segmentation technique to isolate affected lesions from healthy skin. This step is critical as it directly influences the accuracy of disease classification.

3. Feature Extraction:

   • Uses the Gray Level Co-occurrence Matrix (GLCM) to extract texture features from segmented images, which are essential for distinguishing between different skin conditions.

4. Classification:

   • Implements three machine learning algorithms:
     • MobileNetV2: it's the best model so far
     • ResNet50
     • CNN

Classes of Skin Diseases

The paper categorizes skin diseases into eight classes:

1. Melanoma (MEL)
2. Melanocytic Nevus (NV)
3. Basal Cell Carcinoma (BCC)
4. Actinic Keratosis (AK)
5. Benign Keratosis (BKL)
6. Dermatofibroma (DF)
7. Vascular Lesion (VASC)
8. Squamous Cell Carcinoma (SCC)
9. Chickenpox
10. Measles
11. Cowpox
12. Hand-foot-mouth disease or HFMD
13. Healthy ( it's for detecting normal skin)

Datasets Used

The study utilizes two prominent datasets for training and validation:

• MPOX SKIN LESSION 2.0
• HAM10000

Block Diagram of Work

The block diagram illustrates the sequential flow of processes involved in the research methodology:

+---------------------+
|  Image Acquisition  |
+---------------------+
          |
          v
+---------------------+
| Image Preprocessing |
|                     |
|  - Noise Removal    |
+---------------------+
          |
          v
+---------------------+
|    Segmentation     |
+---------------------+
          |
          v
+---------------------+
| Feature Extraction  |
|  - GLCM             |
+---------------------+
          |
          v
+---------------------+
|   Classification    |
|   - MobileNetV2     |
|   - ResNet50        |
|   - CNN             |
+---------------------+

Technology

The proposed system leverages several technologies:

• Image Processing Techniques: For preprocessing and segmentation.
• Deep Learning Algorithms: To classify the segmented images into predefined categories.
• Computational Tools: Likely includes programming environments such as Python with libraries like OpenCV for image processing and scikit-learn for machine learning.

Algorithm

The algorithmic approach can be summarized as follows:

1. Input Image: Acquire an image of the skin lesion.
2. Preprocessing:
   • Convert RGB images to grayscale.
   • Apply Black-Hat transformation followed by inpainting to remove hair.
   • Use Gaussian filtering for noise reduction.
3. Segmentation:
   • Implement GrabCut segmentation to isolate lesions from healthy skin.
4. Feature Extraction:
   • Calculate texture features using GLCM.
5. Classification:
   • Train models using DT, SVM, and KNN on extracted features.
   • Evaluate performance based on accuracy metrics.

This structured approach enhances the reliability of skin disease detection and classification, potentially aiding dermatologists in clinical settings by providing a robust diagnostic tool.