The Impact of the Number of Layers in Convolutional Neural Networks on the Detection of Skin Cancer

Advancements in machine learning have led to important improvements in automated skin cancer detection, with Convolutional Neural Networks (CNNs) emerging as a leading approach. CNNs are widely used in medical imaging due to their ability to extract hierarchical features from images, making them highly effective for classifying dermatological conditions. The architectural complexity of CNNs, particularly the number of layers, is believed to influence accuracy, efficiency and interpretability. Nevertheless, more research is needed to understand this relationship fully. While deeper networks can improve feature extraction, they also introduce challenges such as increased computational demands, risk of overfitting and reduced generalization. Despite the current use of CNNs in medical image analysis, there is limited systematic research on the specific impact of network depth on classification performance. This thesis investigates how the number of layers in a CNN affects its ability to detect and classify skin cancer, aiming to provide insights into optimizing CNN architectures for real-world clinical applications. Through a series of controlled experiments, we evaluate the trade-offs between CNN depth, classification accuracy and computational efficiency using the HAM10000 dataset, a well-established collection of dermatological images. By analyzing performance metrics such as accuracy and F1-score, we aim to determine whether deeper networks consistently lead to better results or if diminishing returns occur beyond a certain depth. Results indicate that classification performance improved with increased CNN depth up to four convolutional layers. The four-layer model achieved the best overall balance between accuracy and generalization, while deeper models showed signs of overfitting and inconsistent gains.