Stomach-Cloud:a Multi-target Trackin Framework For Gastroscodic Video Imag

Gastroscopy examination is one of the most common method for gastric disease diagnosis, computer-assisted gastroscopic images analysis have the potential to improve gastroscopy’s accuracy. Because of the free-form deformation property and the restricted operation space on the gastric internal surface, how to determinate and track the lesion site and no-invasion marking during gastroscopy surgery are still a difficulty and under-consideration issue. In this paper, the research is focused on proposing a multi-target tracking approach under endoscopic conditions.Multi-target tracking in gastroscopy surgery can bring great benefits to many actual endoscopic surgery applications. For example, the physician can mark the suspicious lesion virtually by the multi-target tracking approach at any time during the surgery. Consequently, the virtual marked lesions can be tracked and diagnosed in the following up inspection process. Furthermore, during the process of in-vivo evaluation and therapy planning, the pre-defined virtual marked lesion by the tracking approach can also supply a navigation scheme to the target biopsy site for the physicians, which is helpful to improve the precision of the biopsy targeting and the validity of lesion analysis.Conventional target tracking methods are usually limited to a pair of images with high similarity, and required salient sketches or texture properties. To solve the above problems, this paper proposes a multi-target tracking framework which can be applied to arbitrary target area for image pairs and tracking flexibility. Based on the similarity of the image sequence, the framework proposed in this paper will achieve two advantages:firstly, it can realize target tracking by homography affine theory, which is not depend on salient texture properties in the target tracking objects. Secondly, this approach realizes frame-skip tracking based on a cloud scheme to improve the tracking performance. In this paper, we mainly focus on the following two important technology issues of the multi-target tracking:●The similarity measurement between the image sequences. In this work, we compare five kinds of outstanding image feature detectors, and evaluate the performance comprehensively. Finally, we adopt SIFT algorithm as the measurement of the similarity. The SIFT descriptors are extracted from the gastroscopic image samples, and the matching pairs by the nearest neighborhood algorithm. Moreover, a data cluster method named RANSAC is adopted to reject the false pairs from the matching result. Consequently, the affine transformation is evaluated form the final matching result.●The structure of the stomach cloud. In order to build the gastroscopic images cloud, the similar images are considered as connected nodes in a weighted direction graph, furthermore, the similarity between image sequence is considered as the path weight between corresponding connected nodes. The stomach cloud can supply an effective scheme for calculating the deformation between any image samples in the direction graph. Based on the above assumptions, the reference image samples and target image samples are matched in the stomach cloud, and the deformation relationships are calculated by the stomach cloud automatically, finally target tracking is accomplished based on the deformation relationships.On the basis of the key technology research mentioned above, this paper proposes the multi-target tracking scheme for gastroscopic image sequence, and realize multi-target tracking in gastroscopic video playback process preliminarily. The experimental result shows the gastroscopic image’s tracking accuracy in the offline environment is acceptable, and can supply an effective clinical decision assist for the physicians.