The objective of this paper is to investigate the feasibility of utilizing deep learning-based boundary extraction and semantic segmentation models to achieve an automated and generalized Scan-to-BIM process. Four different input candidates were generated from semantic segments to determine the optimal input for boundary extraction. Results showed that using semantic segments without density information on the state-of-the-art boundary extraction algorithm, Holistic Edge Attention Transformer (HEAT) achieved the best performance with an F1 score of 0.740 for the predicted edges. Further post-processing steps were employed to refine the edges and generate a complete boundary, resulting in an average Intersection-over-Union of 93.5% for 13 rooms in the Stanford 3D Indoor Scene Dataset (S3DIS). These findings demonstrate the potential of leveraging deep learning algorithms for fully automated Scan-to-BIM processes.