Key points are not available for this paper at this time.
Actinic keratosis (AK) is a pre-cancerous skin lesion typically caused by excessive exposure to ultraviolet light. Optical coherence tomography (OCT) can provide sub-surface information relevant for lesion classification, but manual analysis of images is time-consuming and subject to inter-observer variability; automated segmentation based on deep learning models can provide faster and more consistent results. However, structural changes, such as thickening and hyperkeratosis, complicate epidermal segmentation tasks. For this reason, we aimed to develop and optimize a U-Net model for the automated epidermal segmentation of AK lesions in OCT images, combining both accuracy and computational efficiency. Multiple configurations were evaluated by varying hyperparameters, including image sizes (256 × 256, 512 × 512, 1024 × 1024, and 464 × 1356 pixels), batch sizes (2, 4, 8, and 16), and training durations (50, 100, and 150 epochs). Model performance was evaluated quantitatively using several metrics including the Dice coefficient and Jaccard index, comparing automated epidermal segmentations against expert annotations. The optimal configuration, with an image resolution of 256 × 256 pixels and a batch size of 2 over 50 epochs, had a Dice score of 0.86 and a Jaccard index of 0.76. Higher resolutions and longer training increased computation without significantly enhancing the accuracy values and sometimes caused overfitting. These findings show that a simple U-Net architecture could achieve efficient and accurate epidermal segmentation in AK lesions, provided it is fine-tuned to enhance performance.
Angelis et al. (Fri,) studied this question.