*Artificial Intelligence in Surgery, Surgical Subspecialties, and Related Disciplines DOI: http://dx.doi.org/10.5772/intechopen.112691*

autonomy and is rather completely controlled by the operator, a 1 represents a robot that the operator controls but provides some degree of assistance, and 2–5 represent varying levels of autonomy; a 5 represents "true autonomy" of the machine without need for human intervention [109]. Currently, most surgical machines score at level 0 or 1, with machines such as the da Vinci surgical system and robotic endoscopic systems falling squarely in these categories [110]. Applications of level 2 automated robots, such as performing autonomous suturing, have been described [111]. At the current stage, automatons are limited to the autonomy of simple tasks, though there is a push to develop machines that may autonomously perform more complex tasks. Some experiments using phantom tissue have shown success using autonomous robots to ablate abnormal tissue or perform anastomosis of the small bowel, but these experiments were performed on phantom tissue in idealized experimental settings with low trial numbers [112]. Still, these proof-of-concept experiments show that higher-level autonomous robots might emerge sooner rather than later. These complex autonomous robots would integrate multiple sensory modalities, from computer vision to tactile sensation to proprioceptive or auditory information [113].

As AI gets more complicated, the process of training also becomes increasingly complex. Three main learning methods exist for visual-based learning for artificial intelligence: imitation learning, reinforcement learning, and transfer learning [114]. Imitation learning is a method of learning involving the observation of an expert performing the task. Based on the observed actions, the algorithm updates its knowledge (also known as policy) to be more like the demonstration [115]. In an ideal environment, imitation learning will lead to the most reproducible behavior [116]. The use of imitation learning in surgery is limited because of its inability to generalize behaviors. When environments are dissimilar to the demonstration environments, such as differing orientation of visceral organs or working with anatomical variations, the performance of imitation learning algorithms will be suboptimal [116]. This can be alleviated somewhat by dividing the imitation task into subtasks and training subtasks depending on starting circumstances. However, generalizability is still lower than in the other learning methods [115].

Reinforcement learning is another type of learning that is used in AI. This method of learning involves trial-and-error, where the agent performs its task and updates its actions based on the outcomes of its actions. An example of reinforcement learning is the training of the chess engine AlphaZero, in which the engine played many simulated games with itself and improved its playing ability based on the outcomes of each game [117]. Reinforcement learning is a powerful tool that is better able to generalize behaviors compared to imitation algorithms, but reinforcement requires many trials to optimize performance. Additionally, training a model in a real surgical environment is dangerous.

Fortunately, AI flaws can be circumvented via transfer learning, which essentially involves the agent learning through reinforcement learning in a simulated environment and transferring its knowledge to a real environment [114]. Using the simulation, the agent can quickly be trained on many trials before being transferred to real circumstances. Issues for transfer learning are readily apparent; when there is discordance between simulation and real environments, the performance of the model will be suboptimal. A few methods have been proposed to improve transfer learning outcomes. One method is simply improving the quality of the simulation. Computational simulations are much more efficient than physical manipulations of simulated environments, and improvements in computational power are enhancing virtual simulation environments to better model the real world. Other methods involve changing

the policies of the agents to better adapt to circumstances that were not seen during simulation training. One proposed system involves the learning of multiple skill latents in simulation. Broadly defined, "skill latents" represent prelearned or predetermined "primitive skills" which can be subsequently combined within a "model-predictive control" environment to perform more complex tasks [118]. These skill latents can then be accessed and simulated in real time when situations arise that have not been seen before, and the skill latents that produce the optimal effect can be chosen for the agent's actions [118]. Instead of perfectly modeling the real world, this approach tries to make the AI's learning as flexible as possible and/or applicable. Because transfer learning models can be trained in simulation, and because these models can be adaptive, it is likely that autonomous surgical robots in the near future will use transfer learning models to navigate the surgical field (**Table 5**).


*Artificial Intelligence in Surgery, Surgical Subspecialties, and Related Disciplines DOI: http://dx.doi.org/10.5772/intechopen.112691*


**Table 5.**

*Summary of included studies on autonomous robots.*
