Skin Cancer Detection Using CNN with Texture and Feature Analysis
Keywords:
Skin cancer; Melanoma; Dermoscopy; Convolutional neural network; Texture analysis; GLCM; Local binary patterns; ABCD rule; Feature fusion; Attention mechanismAbstract
Early and accurate diagnosis of skin cancer significantly improves patient outcomes, yet access to trained
dermatologists remains limited in many institutional and community healthcare settings, and manual visual
inspection of dermoscopic images is prone to inter-observer variability. Existing automated approaches often rely
solely on deep convolutional features and under-use the complementary texture and handcrafted information that
dermatologists themselves rely on. This paper presents TexFuse-Net, a lightweight skin cancer detection
framework that combines convolutional neural network (CNN) feature learning with texture and handcrafted
feature analysis to classify dermoscopic lesion images as benign or malignant. Images are first cleaned by
morphological hair removal and Shades-of-Gray colour constancy, and the lesion is segmented with a U-Net.
Three feature streams are then computed: deep features from a fine-tuned EfficientNet-B0, texture features from
grey-level co-occurrence matrices (GLCM), multi-scale local binary patterns (LBP) and colour moments, and
clinically motivated ABCD features (asymmetry, border, colour and diameter). An attention-based fusion layer
learns how much weight to give each stream, and the model is trained with focal loss to address class imbalance.
The system returns a calibrated malignancy probability, a Monte Carlo dropout uncertainty estimate and a Grad
CAM heat-map as a confidence report. On the HAM10000 dataset, the framework is expected to achieve 93.3%
accuracy, 84.0% malignant-class recall, 83.0% F1-score and an AUC of 0.961, outperforming CNN-only and
handcrafted-only baselines










