Detection of articulated instruments in retinal microsurgery

Instrument detection in retinal microsurgery is still one of the most challenging operations due to illumination changes, fast motion, cluttered background and deformable shape of the instrument. In this work, a new technique is proposed to detect an articulated forceps instrument by modeling it using Conditional Random Field (CRF). The unary potentials of the CRF, which represent the instrument parts, are detected using the deep convolutional neural network, where two probability distribution maps for both the forceps center and its shaft are estimated. The pairwise potentials are modeled using a regression random forest to learn the relation between the instrument parts based on their joint structural features. Sampled combinations from both unary distributions are selected, and each is tested using the regression forest to compute its similarity to the medical instrument structure. The best combination candidate chosen by the CRF predicts the forceps center point (instrument joint point) and the orientation of its shaft. The approach shows high detection accuracy on public datasets and real videos for retinal microsurgery operations.