Hand dermatitis is a prevalent inflammatory skin condition that affects 10–15% of the general populace, often posing diagnostic challenges due to its varied causes and presentations, especially in the early diagnostic stages. This work utilizes a U-Net architecture for automated hand dermatitis detection and segmentation. Our dataset comprises 700 images of diseased hands and an equal number of non-diseased hands, carefully curated from various sources. Data augmentation techniques enhanced dataset diversity for diseased hands, especially to combat scarce data issues. The U-Net model was trained for 20 epochs by leveraging both binary cross-entropy and Dice loss together, achieving a training accuracy of 99.59% and a validation accuracy of 99.12%. The model demonstrated a Dice coefficient of 0.85 and an Intersection over Union (IoU) of 0.78, indicating strong segmentation performance. These results highlight the model's potential in aiding early diagnosis and distinguishing between diseased and non-diseased hands. Despite data limitations, the model's promising results suggest avenues for future improvements, including more extensive and more diverse datasets and exploring other skin conditions for differential diagnosis.