Traditional fingerprint recognition systems rely on external fingerprints, which face significant limitations due to environmental factors like scratches, water stains, and wear that degrade image quality and recognition performance. Additionally, external fingerprints are vulnerable to forgery using materials like silicone, leading to security concerns. In contrast, internal fingerprints, which are the source of external fingerprints, are more resistant to interference, less affected by external damage, and much harder to forge due to their location beneath the skin, offering enhanced security for fingerprint recognition systems. Optical Coherence Tomorgraphy (OCT) has been applied for fingertip volume data acquisition, which contains internal and external fingerprints. Currently, OCT-based internal and external fingerprint extraction algorithms mainly fall into two categories: the En-Face method and the tissue contour extraction method. The En-Face method accumulates the grayscale information of the stratum corneum layer and viable epidermal layer regions in the OCT axial slice (B-scan) images, generating internal and external fingerprint images based on the grayscale differences within the layer region. However, this type of methods are inevitably affected by noise and individual differences, thus affecting the quality and performance of the generated fingerprints. Besides, it offers fast generation speed without requiring complex preprocessing and postprocessing. On the other hand, the tissue contour extraction method determines the positions of the the stratum corneum layer contour and viable epidermal layer contour in the B-scan images and utilizes the grayscale differences around the contour to generate internal and external fingerprint images. Its advantage lies in its ability to generate high-quality fingerprints with good performance if the contours of the stratum corneum layer and viable epidermal layer are accurately fitted. However, it heavily depends on the accuracy of contour extraction, and inaccurate extraction may lead to incomplete fingerprint ridge and valley information, thereby affecting fingerprint quality and performance, and requiring complex preprocessing and postprocessing steps. Combining the ideas of the En-Face method and contour method, this paper proposes a method for OCT internal and external fingerprint extraction based on Soft Label and Coordinate attention U-Net, referred to as the SLCA-UNet method. Firstly, during the annotation process, based on manually annotated stratum corneum and viable epidermal layer contour position, a polynomial curve traversing the contour is fitted. Using this curve as a reference, Gaussian functions are employed to assign different weights to pixels around the curve, ensuring that pixels closer to the contour center have higher probability weights. After annotating the data, a U-Net network and KL loss function are utilized to learn this probability distribution. Secondly, this paper incorporates a coordinate attention module into the U-Net network structure, enabling it to focus on global features of B-scan images without introducing excessive time overhead. This module helps to focus on the positions of the stratum corneum layer region and viable epidermal layer region, better allocating different probabilities to pixels in each tissue layer. Finally, based on the probability distribution results predicted by the network and the grayscale information of the original image, internal and external fingerprint images are generated. This method not only considers the weights of the contour center and surrounding pixels but also reduces the influence of noise and errors, thereby reducing the algorithm's dependency on contour accuracy. Through a series of experiments, this paper demonstrates that the proposed method can indeed improve the quality and performance of generated fingerprints. Experimental results show that our method can extract internal and external fingerprint images from 3D fingerprint volume data of size 1 400×1 800×500 within 140 seconds. Moreover, in a database of 1 280 fingerprint groups, the mean NFIQ 2.0 score for external fingerprints is 35, with an EER matching score of 0.8%, while the mean NFIQ 2.0 score for internal fingerprints is 47, with an EER matching score of 0.8%. The proposed method not only has a short processing time but also exhibits excellent performance and quality.