Psoriasis detection and classification using MobileNetV3 and Coordinate attention

Technical Summary: Psoriasis Detection and Classification using MobileNetV3 and Coordinate Attention

Problem Statement
Psoriasis is a persistent dermatological condition affecting 2–3% of the global population, characterized by red plaques and silvery scales. While not life-threatening, it causes significant psychological distress and can lead to comorbidities like psoriatic arthritis. Diagnosis is challenging due to clinical heterogeneity (seven distinct variants), similarity to other skin conditions, and the subjectivity of manual clinical assessment. Furthermore, existing severity indices (e.g., PASI) are subjective and often fail to account for specific factors like pruritus or difficult-to-treat areas. While Deep Learning (DL) has shown promise in medical imaging, there is a lack of comprehensive datasets for multi-class psoriasis classification, and existing models often struggle with class imbalance and subtle inter-class variations in skin lesions.

Methodology
The study proposes a deep learning framework utilizing MobileNetV3-Small integrated with a Coordinate Attention (CA) mechanism to classify psoriasis images into seven distinct variants. The methodology follows a structured pipeline:

1. Dataset Construction: A custom dataset comprising 2,101 images was created, covering seven variants: Erythrodermic, Guttate, Inverse, Nail, Plaque, Psoriatic Arthritis (PsA), and Pustular psoriasis. The data was sourced from public repositories (Kaggle) and underwent preprocessing, including resizing to 96×96 pixels and normalization of pixel intensities to the range [-1, 1].
2. Model Architecture:
   • Backbone: The study employs MobileNetV3-Small, chosen for its low computational cost and latency while maintaining accuracy. It utilizes depthwise separable convolutions.
   • Attention Mechanism: To address the limitation of standard Squeeze-and-Excitation blocks (which lack positional information), the authors integrated Coordinate Attention (CA). This mechanism performs horizontal and vertical pooling to encode spatial information, generating attention maps that guide the model to focus on specific "where" and "what" features within the lesion.
   • Classification Head: The architecture includes a Global Average Pooling layer, a Dense layer with 512 neurons using Swish activation, and a Softmax output layer for the seven classes.
3. Training Strategy:
   • Transfer Learning: The model was initialized with weights pre-trained on ImageNet. The top layers were unfrozen for fine-tuning starting from layer 150 to extract high-level features specific to the psoriasis dataset.
   • Loss Functions: The study compares Cross-Entropy Loss (CL) against a proposed Hybrid Loss (HL). The HL combines Cross-Entropy, Focal Loss, and Label Smoothing techniques to mitigate class imbalance and overfitting. The composite loss is defined as $L_{Composite} = \beta L_{CL} + \phi L_{FL}$, where hyperparameters were tuned (specifically $\phi = 0.75$).
   • Optimization: The Adam optimizer was used with a learning rate of $1 \times 10^{-5}$ over 50 epochs.

Key Contributions

• Dataset Creation: Development of a comprehensive dataset containing 2,101 images across seven psoriasis variants, addressing the scarcity of public multi-class psoriasis data.
• Architectural Innovation: Demonstration that integrating Coordinate Attention into MobileNetV3 significantly enhances the detection of subtle variations among the seven psoriasis subtypes.
• Loss Function Optimization: Introduction and validation of a Hybrid Loss function (combining CL, Focal Loss, and Label Smoothing) which outperforms standard Cross-Entropy loss in handling class imbalance and improving accuracy.
• Comparative Analysis: A rigorous comparison against state-of-the-art models, including VGG19, an Ensemble of ResNet models (ResNet50V2, ResNet101V2, ResNet152V2), and standard MobileNetV3, under identical experimental conditions.

Results
The proposed MobileNetV3-Small + CA + Hybrid Loss model achieved the highest performance metrics:

• Accuracy: The model reached 98.40% accuracy on the multi-class classification task. This represents a significant improvement over the MobileNetV3-Small with CA using only Cross-Entropy loss (93.20%) and other baseline models (VGG19: 94.02%, Ensemble ResNet: 91.20%).
• Convergence: Training curves indicated that the proposed model achieved rapid convergence with high training accuracy (97–98%) and low training loss (0.25).
• Generalization: The model demonstrated superior validation accuracy (approx. 75–78%) and lower validation loss compared to models using only Cross-Entropy, indicating reduced overfitting.
• Class-wise Performance: The model showed improved sensitivity for challenging classes such as Inverse and Nail psoriasis, which often suffer from misclassification in standard CNNs.

Significance and Claims
The paper claims that the integration of Coordinate Attention and Hybrid Loss provides a reliable, low-cost support system for dermatologists. By achieving high accuracy in distinguishing between seven specific variants, the model addresses the limitations of manual diagnosis, such as interpretive variability and time consumption. The authors assert that this approach offers precise technical support for managing large volumes of skin images, potentially improving patient quality of life through earlier and more accurate diagnosis.

Limitations and Future Scope
The authors acknowledge several limitations:

• Data Scarcity: The dataset, while improved, remains limited in size and diversity regarding lighting conditions and disease traits.
• Segmentation: The current work focuses on classification; future research must integrate region-of-interest segmentation to handle real-world scenarios where lesions are not perfectly isolated.
• Computational Efficiency: While MobileNetV3 is lightweight, further reduction in computational costs is needed for deployment.
• Correlation Challenges: The study notes that high correlation between certain psoriasis images can still lead to misidentification, suggesting a need for further investigation into distinguishing highly similar variants.