A Deep Learning Framework for Accurate Skin Disease Diagnosis Using CNNs
Dr. Mungara Kiran Kumar,
T Sri Venkata Narayana,
T Yaswi Satya Akshith,
R Srivatsan
The proposed project is an automated skin disease diagnosis framework based on deep learning
consisting of convolutional neural networks implemented in a web-based application. The system
will have two different diagnostic functions, which include the classification of the skin disease and
identifying the type of skin cancer. The framework applies the transfer learning based on the pretrained
convolutional neural network models, with a ResNet50-based model applied to general skin disease
and DenseNet201-based model applied to skin cancer. Both models accept the dermal images resized to
224 x 224 pixels and then they do multiclass classification with the help of the softmax activation. The
training plan involves extraction of features and refinement of the deeper layers to promote the domainspecific
learning of features. Regularization methods like dropout and batch normalization are also
added in order to enhance generalization and minimize overfitting. The trained models are deployed
in a system that is built on Flask, which allows authentication of the users, uploading of images and
real time inference. The prediction consists of class label predictions and confidence scores and class
probability distributions. The dual-model architecture enables the system to respond to the different
degrees of diagnostic complexity and at the same time be computationally efficient. The framework
offers an organized system of applying deep learning models to practical healthcare assistance systems
that will allow the opportunity to conduct the initial screening of skin conditions in an accessible and
automatic manner.
