Develop a deep learning model for skin disease detection and classification
The accurate and timely diagnosis of skin diseases is crucial for effective patient treatment,
yet... See more
The accurate and timely diagnosis of skin diseases is crucial for effective patient treatment,
yet existing diagnostic methods are often limited by their subjective nature, time
consuming processes, and a global shortage of dermatologists. While Deep Learning (DL)
models, particularly Convolutional Neural Networks (CNNs), have emerged as a promising
solution for automated skin disease classification, their performance can be impeded by
challenges such as limited dataset size, severe class imbalance, and the risk of overfitting.
Moreover, the visual similarity of lesions to non-pathological structures presents a
significant challenge to accurate distinction.
This research addresses these limitations by proposing an end-to-end framework for
the diagnosis and classification of seven common skin diseases. The proposed method
integrates the state-of-the-art ConvNeXt-Tiny network with a comprehensive two-phase
hybrid data augmentation pipeline. This pipeline, which includes both offline
transformations and online probabilistic augmentations like Mixup and CutMix, is
precisely designed to regularize the model and enhance its ability to generalize from a
diverse range of visual data. After thorough data preprocessing, which included resizing
and normalization, the ConvNeXt-Tiny model was trained using an AdamW optimizer, a
one-cycle learning rate scheduler, and a smoothed Categorical Cross-Entropy loss function.
The method's efficacy was rigorously evaluated using a broad suite of performance
metrics. The experimental results indicate that the proposed approach obtained enhanced
performance, with a classification of 95.21% accuracy, 93.89% precision, 89.87% recall,
98.62% specificity, 91.60% F1-score, and 98.98% AUC. These findings highlight the
model's exceptional strength in discriminating between a wide variety of skin disorders and
its ability to generalize effectively across diverse conditions. Furthermore, a detailed
evaluation of base, quantized, and pruned models demonstrated that dynamic quantization
reduced the model size by 93.1% (from 106.15 MB to 7.28 MB) and decreased the number
of parameters from 27,825,511 to 1,908,480 with only a minor accuracy drop of 0.004,
while magnitude pruning with 30% sparsity accelerated GPU inference by 1.53× (reducing
batch processing time from 642.27 ms to 420.90 ms) with a slight accuracy improvement
of 0.002, establishing a practical balance between high-performance classification and
resource-efficient deployment.
By providing a scalable, consistent, and objective solution for automated skin
disease classification, this research contributes a powerful tool for Computer-Aided
Diagnosis (CAD) systems, which can assist clinicians and ultimately help mitigate the
diagnostic challenges in dermatology.
2025
