Multiple Detection and Diagnosis of Skin Diseases using CNN

Abstract

One of the most sophisticated deep learning techniques, convolutional neural networks (CNNs), has revolutionised
the field of medical image analysis by enabling previously
unheard-of levels of efficiency and accuracy in the identification
of illnesses. Using a heterogeneous dataset containing images of
common skin conditions, including acne, actinic keratosis, basal
cell carcinoma, and melanoma, we examine how well different
CNN models detect and distinguish between these conditions.
The study covers preprocessing techniques like data augmentation and normalisation to increase the resilience of the
models. Furthermore, we investigate the effects of various CNN
architectures (e.g., VGG, ResNet, DenseNet) on classification
performance considering computational efficiency and model
complexity. Through extensive testing and evaluation, we quantify
each model’s classification accuracy, sensitivity, specificity, and
computational overhead, providing insightful data on how wellsuited it is for real-world clinical applications. We use CNN
models to demonstrate the interpretability of the learned features,
assisting dermatologists in understanding the discriminative patterns utilised in disease diagnosis. The proposed framework not
only improves the accuracy and efficacy of diagnosis, but it also
serves as a helpful educational tool that helps physicians better
understand the wide range of skin conditions they treat. When
combined, these results provide credence to the ongoing efforts
to incorporate deep learning into automated health systems,
which could lead to improved patient care and dermatological
diagnostics. To tackle this, deep learning algorithms for accurate
disease classification become crucial. Several machine learning
algorithms were tested, and we found that the Deep CNN model
with contrast stretching yields the best results. The dataset was
divided into two categories: test and train image datasets for
the four diseases mentioned above, and a validation dataset.
Subsequently, we trained our model on this data, yielding an
accuracy of 92.7%.