Intelligent Vitamin Deficiency Detection Through Dermatological Image Analysis using CNN and ViT

Vitamin deficiencies often manifest through visible dermatological changes in multiple anatomical regions such as the tongue, nails, eyes, lips, and skin, enabling non-invasive image-based screening. This paper presents a deep learning–based framework for vitamin deficiency detection using a comparative analysis of Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) architectures. The proposed system processes dermatological images through standardized preprocessing and extracts features using both CNN-based local pattern learning and ViT-based global self-attention mechanisms. An ensemble decision strategy is employed to enhance prediction reliability. Experimental evaluation was conducted on images from five body regions. The proposed system successfully identified Vitamin D deficiency from nail images with an ensemble confidence of 45.0% (CNN: 45.4%, ViT: 44.7%), exceeding the threshold of 42.0%, and classified it as high severity. For eye images, no deficiency was detected, achieving an ensemble confidence of 55.0% (CNN: 46.9%, ViT: 60.4%) above the healthy threshold of 50.0%. Lip image analysis detected Vitamin C deficiency with 47.8% ensemble confidence (CNN: 55.8%, ViT: 42.5%) at a threshold of 45.0%, indicating moderate severity. Tongue images revealed Vitamin B deficiency with an ensemble confidence of 53.0% (CNN: 50.3%, ViT: 54.8%) against a threshold of 40.0%, while skin images identified Vitamin A deficiency with 58.3% ensemble confidence (CNN: 50.2%, ViT: 63.8%) at a 45.0% threshold. The results demonstrate that CNNs perform effectively for localized dermatological features, whereas Vision Transformers exhibit superior performance in capturing global patterns across complex regions such as skin and eyes. The ensemble-based multi-region framework improves diagnostic robustness and provides a scalable, non-invasive solution for early vitamin deficiency screening and clinical decision support.