ΑΚtransU-Net: Transformer-Equipped U-Net Model for Im-Proved Actinic Keratosis Detection in Clinical Photography

The integration of artificial intelligence into clinical photography holds significant poten-tial for enhancing the monitoring of skin conditions, such as actinic keratosis (AK), and the broader phenomenon of skin field cancerization. Accurate identification of AK burden within areas of field cancerization often depends more on contextual cues—such as the surrounding photodamage—than on lesion morphology alone. This reliance on broader spatial context highlights the need for models that can effectively combine fine-grained local features with a comprehensive global view in real-world clinical imaging settings. To address this challenge, we propose AKTransU-net, a hybrid U-Net-based architecture designed to enhance both spatial detail preservation and global contextual understand-ing. The model incorporates ViT-based Transformer blocks at multiple encoding levels to enrich feature representations, which are then passed through ConvLSTM modules em-bedded within the skip connections. This configuration allows the network to maintain semantic coherence and spatial continuity throughout the segmentation process. Such global awareness proves especially critical when applying the model to whole-image de-tection via tile-based processing, where continuity across tile boundaries is essential for accurate and reliable lesion segmentation. The effectiveness of AKTransU-net was demonstrated through comparative evaluations with state-of-the-art semantic segmentation models, showing notable improvements in AK segmentation accuracy.