RNN-LSTM based Tissue Classification in Robotic System for Breast Biopsy

Objective: Accurate needle tip placement for biopsy procedures is difficult to achieve with low-fidelity imaging systems. Conventionally, surgeons while performing biopsies rely on ultrasound images and intuitive feeling about needle-tissue interaction forces to confirm target location. Currently, robotic assistance for biopsy uses only the position parameter to address localization challenges. In the present work, in addition to a robot's position sense, we propose to integrate needle-tissue force parameter. This force model presents a new way to built an intelligent robot that can identify tissue properties during needle probing/biopsy. Methods: A standard experiment was setup that consist of a force sensor, a linear stage, a biopsy needle, and synthetic tissues. During the experiment, needle penetrates through synthetic tissues, a set of data (force and distance) was acquired and manually labeled. A recurrent neural network (RNN) based Long-Short Term Memory (LSTM) model was trained with the data to estimate the various classes (air, skin/fibrous tissue, puncture, and hard tissue). Result: The trained model is able to distinguish between the three synthetic materials. Intuitively, this model mimics human perceptions of force during a handheld needle penetration. Conclusion: The constrained experimental setup helps us present a proof of concept for using deep learning models for tissue classification. Significance: Tissue classification is the first step towards solving the more difficult problem of developing a robotic device capable of precise event detection of tissue transitions.