RM-SEAGR Net: Interventional Control of Retinal Microsurgery Using SEAS Network Segmentation and Modified GRU Model

Retinal microsurgery is extremely difficult and requires highly skilled surgery with very little margin for error in manipulation. Unexpected manipulations could cause the tool and sclera to come into high force contact, which would harm the sclera. When a surgeon performs surgery, the usage of robotic aids may enhance and increase their manipulating power. However current robotic systems still struggle to identify and avoid the potential intraoperative risk posed by a surgeon’s errors. So, in the proposed work, a novel modified deep learning-based interventional control of robot aided retinal surgery is developed. Structured analysis of the retina (STARE) images gathered from the microscope is used as the dataset. These collected raw data may consist of poor visual quality, hence it was enhanced using resize, convolution filtration and regularization reconstruction algorithm. After preprocessing, the enhanced retinal images were segmented using the supervised edge attention guidance segmentation (SEAS) algorithm, which segments the eye nerves and unhealthy portions preciously. The segmented images are extracted using features using the depth wise dilated convolutions and feature reuse-residual block (DDC-FRB) model. These features were given to the modified gated recurrent model (GRU) model for the prediction of healthy and nonhealthy eyes. In GRU, the sigmoid function is modified as well and the performance of the prediction process is improved by selecting the optimal parameter of GRU using the walrus optimization algorithm (WaOA). The proposed interventional controlling system offers 96.39 \(\%\) recall, 3.6 \(\%\) FNR, 91.4 \(\%\) kappa, and 98.2 \(\%\) accuracy. Thus, the proposed modified deep learning classifier and SEAS segmentation model effectively detect the unhealthy portion with more accuracy. This deep learning based model can be used in real-time applications such as health care management to detect heart, kidney, lung and eye disease by training the model with different datasets. In other applications, the trained model can also be implemented in the robotic surgical tool for operating the patient’s organ based on the detection of disease location.