Surgical tool tracking based on two CNNs: from coarse to fine
Zijian Zhao, Sandrine Voros, Zhaorui Chen, Xiaolin Cheng · The Journal of Engineering · 2019
The current convolution neural network (CNN)‐based methods suffer from mass computation and low frame rate without special acceleration, they cannot be applied in some real applications. The authors propose a coarse to fine method for surgical tool tracking based on CNNs. Two CNNs are designed for the coarse and fine locations of the surgical tool. The coarse CNN is a classification network of 10 classes, and the fine CNN is a regression network for the tool tip area. The spatial and temporal context updating makes two CNNs cooperating together for the whole tracking process of surgical tool. The authors validate their method in the experiments with eight datasets, where there are two ex vivo datasets and six in vivo datasets, as well as comparing their method with other four methods in terms of accuracy and speed of tracking. The ex vivo and in vivo experiments demonstrate that the method gives consideration to both accuracy and speed, and provides a good accuracy and a high frame rate of tracking. For the future research of multi‐tool tracking, the design of new regression CNN with the output of tensor will be focused, which contains multiple tools' labels and tracking information.