| Abstract | Background: Skin disorders are among the most prevalent medical
problems in the world, and in order to prevent consequences, a prompt and
precise diagnosis is typically required. However, some of the main obstacles to
automated classification techniques are the lack of annotated clinical images,
overlapping symptoms, and the visual similarity of lesions.
Objective: By combining the relative benefits of DenseNet121 and
EfficientNetB0, the suggested study suggests a Dual-Backbone Transfer
Learning Network (DBTLN) designed to improve the diagnostic performance
of skin disease classification.
Methods: The DBTLN structure makes use of EfficientNet’s depth scaling for
computational efficiency and DenseNet’s dense connectivity of features for
preserving fine-grained lesion features. 19,171 dermoscopic photos of 19
distinct skin diseases were used to train and test the model. Presentation was
assessed using exactness, accuracy, memory, and F1-score, and comparisons
were made with the traditional CNN models, VGG19, MobileNetV2, AlexNet,
DenseNet121, and EfficientNetB0.
Findings: The DBTLN outperformed all baselines with validation accuracy of
97.57%, correctness of 0.95, remembrance of 0.96, and F1-score of 0.95. These
results demonstrate enhanced generalization across a broad range of skin lesion
types, particularly under constrained and irregular clinical settings. Thus, this
model substantially enhances skin disease prediction by leveraging deep feature
fusion and adaptive learning mechanisms. It also achieves reliability by
minimizing prediction error and enhancing generalizability.
|
|---|
