Dual-Architecture Neural Networks for Skin Cancer Lesion Segmentation: A Comparative Study of SingleNet and DoubleNet

Automated segmentation of skin cancer lesions plays a pivotal role in computer-aided dermatological diagnosis by assisting clinicians in identifying malignant regions at early stages, thereby improving treatment planning and patient survival rates. Accurate boundary delineation reduces inter-observer variability and enhances consistency in quantitative analysis, making deep learning–based segmentation systems increasingly valuable in clinical decision support environments. This paper presents a comprehensive study and practical implementation of two convolutional neural network architectures: SingleNet and DoubleNet. SingleNet serves as a structured encoder–decoder baseline framework designed to capture hierarchical image representations through progressive downsampling and upsampling operations. It establishes a reliable reference model for evaluating architectural improvements. DoubleNet extends this design into a dual-stage architecture that processes complementary feature representations sequentially, enabling enhanced contextual awareness and refined boundary detection. The dual-stage processing facilitates richer feature fusion and improved discrimination between lesion and healthy tissue. Both architectures were trained and evaluated using curated dermoscopic imaging datasets. The implementation incorporates stacked convolutional blocks with batch normalization to stabilize gradient flow and accelerate convergence. Nonlinear activation functions enhance model expressiveness, while skip connections between encoder and decoder stages preserve fine-grained spatial information that is typically lost during pooling operations. This design enables effective learning of both local texture patterns and global contextual structures. Performance evaluation was conducted using established segmentation metrics, including Intersection over Union (IoU) and the Dice coefficient, along with training and validation loss curves to assess convergence behavior and generalization capability.