A Deep CNN for Automated Skin Disease Classification

Syed Mehsher, Mehdi Hussain Baba, Sujal Chaudhary, Sanjiv Kumar Tomar

Department of Computer Science, ASET, Amity University, Noida, Uttar Pradesh, India

Abstract in size, lighting, and quality, which can affect model (b) Threshold-based decision mechanism applied on
accuracy if not standardized. Preprocessing techniques such feature maps to improve classification confidence and final
as resizing, normalization, and basic filtering are applied to prediction.
Recent advancements in artificial intelligence, computer prepare the data. Resizing ensures a uniform input size (e.g.,
vision, and deep learning have significantly enhanced the 224×224), while normalization scales pixel values to a
capabilities of healthcare diagnostic systems. The rapid consistent range, improving training stability and
Experimental Results
evolution of image processing techniques and convolutional convergence. The processed images are then passed through
neural networks (CNNs) has made it feasible to analyze skin convolutional and pooling layers in the CNN, which extract The experiments and validation of the proposed model have
conditions using digital images. meaningful features such as edges, textures, and patterns, been carried out using multiple dermatological image
enabling accurate classification of skin diseases. datasets. Fig. 4 presents the classification results of different
In this work, a skin disease detection system is developed skin diseases using the trained CNN model. Performance
that leverages deep learning models to classify various metrics such as accuracy and loss are evaluated across
dermatological conditions based on image inputs. The system different training epochs. The average performance scores
evaluates performance metrics such as accuracy, precision, are summarized in Table 1 for different disease categories.
recall, and loss to assess classification efficiency.
The results indicate that:
Variations in model performance are analyzed based on
dataset quality, preprocessing techniques, and feature ● Accuracy improves steadily with training
extraction methods. For the considered dataset, higher ● Loss decreases as the model converges
accuracy is observed in well-defined skin conditions, while ● Performance varies depending on disease
comparatively lower performance is noted in visually similar complexity and dataset quality
disease classes. Therefore, a comprehensive analysis of
multiple skin disease categories has been conducted to
optimize classification accuracy and reliability. The results Fig.2: Image Preprocessing and Feature Extraction Pipeline The average performance scores across different disease
indicate progressive improvement in detection accuracy with categories are summarized and compared to evaluate overall
enhanced preprocessing and model tuning, achieving system effectiveness.
The preprocessing techniques are essential when handling
maximum efficiency in clearly distinguishable skin patterns. real-world skin images affected by noise, lighting variations,
and background distortions. With sufficient data
transformations, the model learns robust feature
Proposed Technique representations, improving classification performance across
different skin conditions.
Convolutional Neural Network (CNN): It is considered a
highly effective deep learning approach for image B. Challenges in Classifying Visually Similar Skin
classification tasks, capable of automatically extracting
Diseases:
spatial features and patterns from skin images. The model
provides accurate classification of different skin diseases by
learning hierarchical feature representations. As the visual similarity between different skin diseases
increases, distinguishing between them becomes more
challenging, often resulting in overlapping feature
The model processes dermoscopic images through multiple representations. When the difference between disease Fig.4: Classification results of different skin diseases using
convolutional and pooling layers, enabling it to learn patterns is minimal, the model may produce closely related the trained CNN model.
complex patterns such as edges, textures, and lesion prediction probabilities, leading to classification ambiguity.
structures. These learned representations allow the system to Here, feature maps represent pixel-level information, and
accurately classify different skin diseases with high threshold-based decision mechanisms are applied to improve Conclusions
reliability and consistency. classification confidence.
This often results in:
AI-based healthcare systems provide an efficient and scalable
● Overlapping feature representations in the
solution for early detection of skin diseases. Since such
learned feature space
systems operate under computational constraints and depend
● Reduced classification confidence
on dataset quality, optimizing model performance and
● Higher chances of misclassification
resource utilization becomes essential. The proposed system
demonstrates improved classification accuracy while
maintaining computational efficiency. The results indicate
that the model effectively reduces misclassification and
enhances prediction reliability compared to traditional
approaches. Thus, the developed system contributes to
improving automated dermatological diagnosis and supports
early-stage medical intervention.

Fig.1. Diagram of Proposed Algorithm

A. Image Preprocessing and Data Augmentation: