NJMU Bone Drilling Force Dataset

In order to train a neural network to predict bone drilling force, we build this dataset. The force data in this dataset come from two sources. The first source is the physics model-calculated force data obtained based on physical cutting laws validated by researchers in this field. Since this cutting process can be simulated by programs, the data volume is almost unlimited. The second source is sensor-recorded force data, which reflect the actual bone drill force imposed on the drill bit during operations but are limited by the utilized equipment and the complexity of experiments. Therefore, we generate a large scale synthetic dataset (Synthetic dataset.zip) with program generated images and physics model-calculated force data to pretrain the cutting physics model-guided network. Then, we build a real bone dataset with bovine bone CT images and sensor-recorded force data to fine-tune the previous network and train the PFPP and IRTP network to predict sensor-recorded force values.The synthetic dataset contains 985,356 images from 600 virtual drilling paths. Each path is set with specific drilling conditions parameters (drill_geometry_working_station.mat). Each image is paired with the radial force X, radial force Y, thrust force, and torque calculated by the physics model (data_model_norm/unnorm.csv). The real bone dataset contains 9162 images from 14 drilling paths from 3 bovine bones. Each path is set with specific drilling conditions parameters. Each image is paired with the radial force X, radial force Y, thrust force, and torque calculated by the physics model and thrust force measured by force sensor (data_forcesensor_norm/unnorm.csv).