A real-time neurosurgical instrument segmentation method based on endoscopic images

Abstract

The invention belongs to the field of medical image processing and image segmentation technology, and more particularly, relates to a real-time neurosurgical instrument segmentation method based on endoscopic images. A set of real-time instrument instance segmentation methods for endoscopic neurosurgery scenarios are proposed, which can be applied to clinical practice and play a role in assisting neurosurgery in real time during surgery. The invention also proposes a set of data augmentation methods for light spots, reflections, blurs and other noises, so as to enrich the samples and at the same time improve the learning ability and adaptability of the model.

Description

technical field [0001] The invention belongs to the field of medical image processing and image segmentation technology, and more particularly, relates to a real-time neurosurgical instrument segmentation method based on endoscopic images. Background technique [0002] Existing instance segmentation methods are mainly divided into two types: two-stage and one-stage. At present, there is no relevant real-time instance segmentation work in the neurosurgical endoscopic image scene. [0003] Data augmentation is one of the commonly used techniques in deep learning. It is mainly used to increase the training data set to make the data set as diverse as possible, so that the trained model has stronger generalization ability. Existing data augmentation mainly includes: horizontal / vertical flip, rotation, scaling, cropping, shearing, translation, contrast, color jitter, noise, etc. However, the traditional data augmentation methods are not aimed at endoscopic surgical images, nor a...

AI Technical Summary

Problems solved by technology

It has not been researched and developed for the scene of neurosurgery endoscopic surgery, and cannot solve the challenges caused by noises such as light spots, reflections, and blurs that are prone to appear in this scene

At the same time, most of the above instance segmentation techniques help doctors in the preoperative diagnosis stage, and cannot provide real-time prompts during the operation.

[0005] At present, the instance segmentation algorithms with good effect are all derived from the target detection method, but the instance segmentation is much more difficult than the target detection.

The accuracy of two-stage detectors relies on feature localization, which is sequential and cannot be accelerated

The single-stage detector improves the process into a parallel process, but this requires a lot of subsequent calculations after positioning, and it is difficult to speed up

Therefore, the real-time instance segmentation task has been difficult to break through

Images