With the introduction of robotic colorectal surgery, colorectal cancer surgical treatment has entered a new era. The da Vinci System is now the most extensively utilized robotic surgical. It allows surgeons to execute extremely delicate or highly complex procedures with wristed devices that have seven degrees of freedom. When compared to traditional open surgery, the advantages of surgery with these robots include a shorter period of recovery and hospital stay, minimum scarring, smaller incisions, and a significant reduction in the risk of surgical site infections, postoperative pain, and blood loss (26, 27). Surgeons can operate with a larger viewing field thanks to computer-controlled instruments. Computer-controlled devices enable surgeons to work with a wider viewing field, greater flexibility, dexterity, precision, and less fatigue. The da Vinci dual-console provides integrated teaching and supervision, for residents’ surgical training. The Senhance surgical robot (TransEnterix Surgical Inc., Morrisville, NC, USA) is a laparoscopic-based technology that allows skilled laparoscopic surgeons to perform more sophisticated surgeries (28).

The robotic platform has a distinct benefit in that it allows access to difficult-to-reach locations, such as a narrow pelvis, while also preserving postoperative urinary and sexual function (29). Rectal surgery was related with higher conversion rates in men, obese patients, and patients getting a low anterior resection compared to robotic surgery in the ROLARR randomized clinical study (30). Recent research has also showed that robotic-assisted surgery appears to be more suitable for protecting the pelvic autonomic nerve (29, 31, 32).

The Intuitive Surgical da Vinci system pioneered the notion of transparent teleoperation, in which motions done by the surgeon on the control interface are precisely copied by surgical tools on the patient side. The lack of a decision-making process by the machine in the transparent teleoperation paradigm gives the surgeon unlimited control. However, these devices have certain algorithmic autonomy, such as tremor suppression and redundancy resolution, which do not interfere with the surgeon’s actions (33). They can not be considered AI driven devices, but they offer a starting point for the hardware for future autonomous operating robots.

Shademan et al. described complete in vivo autonomous robotic anastomosis of porcine intestine utilizing the Smart Tissue Autonomous Robot (STAR) (16). STAR surpassed human surgeons in a range of ex vivo and in vivo surgical tasks, despite being conducted in a carefully controlled experimental context. In later in vivo tests, STAR obtained 66.28% correctly placed stitches in the first attempt, which corresponded to an average of 0.34 suture hesitancy per stitch (17).

For the first time, these experiments proved the fledgling clinical feasibility of an autonomous soft-tissue surgical robot. STAR was controlled by artificial intelligence (AI) algorithms and received input from an array of optical and tactile sensors, as opposed to traditional surgical robots, which are managed in real time by people and have become ubiquitous in specific subspecialties.

Phase recognition is the task of identifying surgical images according to preset surgical phases. Phases are parts of surgical operations that are required to finish procedures successfully. They are often determined by consensus and recorded on surgical videos (34).

There are several studies of phase and action recognition, on different types of surgery, including colorectal surgery. The study of Kitaguchi et al. aimed to create a large annotated dataset containing laparoscopic colorectal surgery videos and to evaluate the accuracy of automatic recognition for surgical phase, action, and tool by combining AI with the dataset. They used 300 intraoperative videos and 82 million frames were marked for a phase and action classification task, while 4000 frames were marked for a tool segmentation task. 80% of the frames, were used for the training dataset and 20% for the test dataset. CNN was utilized to analyze the videos. The accuracy for the automatic surgical phase task was 81%, while the accuracy for action classification task was 83.2% (18).

The creation of an image-guided navigation system for areolar tissue in the complete mesorectal excision plane using deep learning has been reported by Igaki et al. This could be useful to surgeons since areolar tissue can be utilized as a landmark for the optimum dissection plane. Deep learning-based semantic segmentation of areolar tissue was conducted in the whole mesorectal excision plane. The deep learning model was trained using intraoperative images of the whole mesorectal excision scene taken from left laparoscopic resection movies. Six hundred annotation images were generated from 32 videos, with 528 photos used in training and 72 images used in testing. The established semantic segmentation model helps in locating and emphasizing the areolar tissue area in the whole mesorectal excision plane (19).

There are commercial systems available that allow the endoscopic camera to move without human intervention, following particular features in the scene., Viki (35), FreeHand (36), SOLOASSIST (37)and AutoLap (20), for example, do camera stabilization and target tracking. These were the first autonomous systems used to assist with MIS intervention. The autonomy is implemented via feature tracking algorithms that maintain the surgical instrument in the endoscope’s visual field (38).

For AutoLap minimally invasive rectal resection with entire mesorectal excision was chosen to experimentally test cognitive camera guidance as this surgical method places great demands on camera control. A single surgeon performed twenty surgeries with human camera guidance for learning purposes. After the completion of the surgeon’s learning curve, two different robots were trained on data from the manual camera guiding, followed by using one robot to train the other. The performance of the cognitive camera robot improved with experience The duration of each surgery improved as the robot became more experienced, also the quality of the camera guidance (evaluated by the surgeon as good/neutral/poor) improved, becoming good in 56.2% of evaluations (20).

In order to predict anastomotic complications attributable to hypoperfusion after laparoscopic colonic surgery, a fluorescence laparoscopic system can be used during surgery for angiography using indocyanine green (ICG). Each patient has a different perfusion status, due to individual variations in collateral circulation blood flow pathways, which provides a different ICG curve. A well-trained AI can forecast the probability of hypoperfusion-related anastomotic problems by analyzing the microcirculation state, by using numerous metrics and ICG curve patterns. The AI-based micro perfusion analysis system can help surgeons by quickly performing real-time analysis and giving information in a color map to surgeons. Using a neural network that imitate the visual cortex, Park et al. clustered 10,000 ICG curves into 25 patterns using unsupervised learning, an AI training approach that does not require annotations during training. ICG curves were derived from 65 processes. Curves were preprocessed to minimize the degradation of the AI model caused by external factors such light source reflection, background, and camera movement. The AI model revealed more accuracy in the microcirculation evaluation when the AUC of the AI-based technique was compared to T1/2 max max (time from first fluorescence increase to half of maximum), TR (time ratio: T1/2 max/Tmax, Tmax is the time form first fluorescence increase to maximum), and RS (rise slope), with values of 0.842, 0.750, 0.734, and 0.677, respectively. This makes it easier to create a color mapping scheme of red-green-blue areas that classifies the degree of vascularization. In comparison to a surgeon’s solely visual inspection, this AI model delivers a more objective and accurate approach of fluorescence signal evaluation. It can provide an immediate evaluation of the grade of perfusion during minimally invasive colorectal procedures, allowing for early detection of insufficient vascularization (21, 39).

Knot-tying is part of basic surgical skills and a quick technique in open surgery, while laparoscopic knot-tying can take up to three minutes for a single knot to be done. Mayer et al. described a solution based on RNNs to speed up knot-tying in robotic cardiac surgery. The surgeon inputs a sequence to the network (for example, instances of human-performed knot-tying), and an RNN with long-term storage learns the task. The preprogrammed controller was able to construct a knot in 33.7 seconds, however the introduction of an RNN offered a speed improvement of about 25% after learning from 50 prior runs, generating a knot in 25.8 seconds (22, 40, 41).

Optical biopsy is a light-based nondestructive in situ assessment of tissue pathologic features. Hyperspectral imaging (HSI) is a non-invasive optical imaging tool that provides pixel-by-pixel spectroscopic and spatial information about the investigated area. Tissue-light interaction produces distinct spectral signatures, allowing the visualization of tissular perfusion and differentiation of tissue types. HSI cameras are commercially available and are easily compatible with laparoscopes (42, 43).. In the past years several very promising studies, which used different AI methods in detecting CRC during surgery using HIS, were published.

Jansen-Winkeln et al. used HSI records from 54 patients who underwent colorectal resections, creating a realistic intraoperative setting for their study. By using a CNN method, they obtained a sensitivity if 86% and specificity of 95% for the distinction between cancer and healthy mucosa, while differentiating cancer against adenoma had a sensitivity of 68%, and 59% specificity o for CCR (23).

Collins et al. used HIS imaging on specimens obtained immediately after extraction from 34 patients undergoing surgical resection for CRC. Using a CNN to automatically detect CRC in the HIS images they obtained a sensitivity of 87% and specificity of 90% for cancer detection. Their approach could be used for objectively assessing tumor margins during surgery (24).

By combining HSI and CNN trained with deep learning on porcine models, Okamoto et al. obtained an automatic distinction of different anatomical layers in CRC surgery, achieving a recognition sensitivity of 79.0 ± 21.0% for the retroperitoneum and 86.0 ± 16.0% for the colon and mesentery (25)..These results are promising in improving the results of complete mesocolic excision, by lowering the complications associated with it (like lesions of the ureter, gonadal vessels)and offering a better oncologic result.