This project employs deep learning techniques to classify skin cancer lesions using CNNs. Through comparative analysis, a CNN achieves 97% accuracy, while a pre-trained ResNet-152V2 model attains 95%, showcasing the efficacy of deep learning in dermatological diagnosis.
In dermatology, where a precise diagnosis is essential for efficient treatment and positive patient outcomes, skin cancer presents a substantial obstacle. The goal of this research is to employ Convolutional Neural Networks (CNNs) to build a deep learning model for the classification of skin cancer lesions. The project uses the HAM10000 dataset, which includes a variety of photos of skin lesions, to perform a thorough exploratory data analysis in order to comprehend the features of the dataset and address concerns with class imbalance by using oversampling techniques.
The main goal of the study is to compare the effectiveness of two different methods: a pre-trained ResNet-152V2 model and a conventional CNN that is trained from scratch. The study seeks to achieve optimal classification accuracy across seven different types of skin cancer lesions by optimizing hyperparameters and fine-tuning the models.
Experimentation phase results show how well deep learning techniques work for correctly classifying skin cancer lesions. With an astounding 97% accuracy rate, the CNN model demonstrates its capacity to identify minute patterns and characteristics that correspond to a range of skin disorders. Furthermore, a competitive accuracy of 95% is obtained by using a pre-trained ResNet-152V2 model that was initialized with weights from ImageNet.
This study shows how deep learning may be used to diagnose dermatological conditions, but it also shows how important it is to use models and prior knowledge to improve classification results. The project advances the field of computer-aided diagnosis in dermatology by means of thorough data analysis, model training, and comparative evaluation. The ultimate goal is to achieve more accurate and efficient detection of skin cancer lesions.