Operation Images Context Recognition Based On Transfer Learning

Along with the progress of medical technology, there is also a sharp increase in medical surgery demand; The type of surgery increase quickly and the procedure of surgery become more and more complex, which increase the difficult of training new doctors; The modern surgeries are more and more dependent on the computer aided surgery system than before. In addition, a surgery need a measurement to measure its rightness. All these facts increase the demand for completely understanding the data in surgery, researchers have proposed Surgery Procedure Modeling (SPM) method to solve this problem. SPM need computer to understand the surgery video in human manner. How to across the semantic gap between low-level features and high-level semantic is the key step of SPM, and it is also a more challenging study subject in video analyses.Surgery videos are quite different with traditional object recognition and video analyses dataset. It is a perfect reproduction of a surgery. In a surgery, everything is possible. And because of the complexity of background and the occlusion of different objects, surgery video analyses is a more challenging task. Recently, the methods of SPM are mainly based on the existed method of computer vision field to analyses the surgery video in a whole, extract the features suit for videos such like optical flow features. But these methods often ignore the gap between low-level features of every frame in videos and the high-level video semantic. For solving this problem, this paper proposed that applying the Convolutional Neural Network (CNN) model to this field, and analyses the objects and its position of every frame in videos with the help of CNN.CNN is the most popular method in image classification field. It is a supervised learning method and learning the hierarchy feature representation automatically. But CNN needs lots of labeled images to avoid over-fitting problem. But there is no existing labeled datasets in medical field, and because of the professional limits of medical, we can't collect a very large labeled datasets in short time. We collect a small datasets of medical instruments from public medical operation videos, and then label it by hand. This paper proposed that applying the transfer learning method of machine learning to solve the problem of lacking data.The experiment result indicated that despite the lacking of labeled data, the complexity of background, the occlusion of different objects, the combination of CNN and transfer learning can also get a good performance. The results also shows strong evidence of object localization, which is import for SPM model.