The practical implementation of maximum likelihood detection is limited by its high complexity as well as requiring perfect channel state information. Although conventional blind detection techniques reduce complexity, they degrade performance and require blind channel state information. In this paper (for the first time), a deep learning based blind detection and a joint blind detection-constellation shaping structure are presented (to solve this problem). This paper (deeply) goes through the problem and discusses several (practical) scenarios, including single user, multiuser with resource (channel) allocation, and multiuser without resource allocation (multiuser interference). In order to show the universality of the proposed systems, wide atmospheric turbulence regimes, from weak to strong are considered, and single input single output, as well as multi-input multi-output structures are considered. Results indicate that without channel estimation, a deep learning based (blind) detector (despite its very few complexity, and despite conventional systems require it), could have a very favorable performance at all around. So, it is expected that practical implementations of the proposed structures greatly reduce cost, and processing latency, while maintaining performance close enough to the outstanding conventional systems.