Text guided cross attentive multimodal learning with visual feature modulation for automated skin lesion detection

Automated skin lesion detection is essential for early dermatological diagnosis. Most deep learning algorithms employ dermoscopic images and ignore the clinical context of dermatologists. This constraint decreases the robustness and interpretability, particularly in visually ambiguous instances. An explainable multimodal cross attention framework that merges clinical text with dermoscopic images to increase diagnostic accuracy and semantic grounding in automated skin lesion detection. A Text-Guided Cross-Attentive Visual Feature Network multimodal learning architecture (TG-CAVNet) using Bio-ClinicalBERT based clinical text encoding and EfficientNet-B4 visual feature extraction is proposed. The framework uses text-guided channel-wise feature modulation, text-queried cross-attention for semantic-spatial alignment and adaptive multi-stream fusion to merge complementary representations. The model was trained end-to-end using hybrid focal and cross-entropy losses. On a multimodal dermoscopic dataset of 6194 aligned image-text samples, TG-CAVNet outperforms state-of-the-art multimodal baselines with 90.75% accuracy and a macro Jaccard score of 0.82. Ablation investigations are performed that confirmed the separate and synergistic impacts of the components, whereas attention visualizations improved interpretability. Text-guided cross-attentive multimodal learning improves the performance and explainability of automated skin lesion identification. The robust and clinically interpretable decision-support framework TG-CAVNet demonstrates the necessity of integrating semantic clinical context with visual analysis in dermatological AI systems.