A MACHINE-LEARNING-ASSISTED TRAJECTORY OPTIMIZATION OF AUTONOMOUS WOUND IMAGING WITH THE ABB IRB 120 SIX-DEGREE-OF-FREEDOM MANIPULATOR

Abstract

The documentation of traumatic wounds by manual methods is attended by inter-operator variability and, on occasion, by the inconvenience of direct contact with the wound site. A robotic manipulator, furnished with a medical-grade optical or depth sensor, would in principle offer a means of acquiring wound images of a standard uniform enough to support the data-driven analytical methods that have lately come to the fore in clinical practice.

A kinematic framework for the ABB IRB 120 six-degree-of-freedom industrial manipulator is here described. The forward kinematic solution is obtained by successive multiplication of Denavit–Hartenberg transformation matrices; the inverse kinematic problem is solved numerically by damped Jacobian pseudo-inverse iteration, with the damping factor chosen to maintain numerical stability in the vicinity of singular configurations. Singularity detection proceeds through evaluation of the Jacobian determinant, and kinematic dexterity is assessed by the Yoshikawa manipulability measure. A corpus of ten thousand simulated wound scenarios was generated within the MATLAB environment, the scenarios varying independently in length, width, depth, orientation, surface curvature, scanning velocity and tool-path strategy. Upon eighty per cent of the cases, five supervised regression models were trained: an artificial neural network, the k-nearest neighbours algorithm, ordinary linear regression, support vector regression with Gaussian kernel, and a random forest ensemble. The remaining two thousand cases were reserved for evaluation.

The linear regression and support vector models proved jointly the most accurate for prediction of trajectory efficiency, each attaining a coefficient of determination R² = 0.997, with root mean square errors of 5.035 per cent and 5.028 per cent respectively. In the prediction of scan time the linear regression led with R² = 0.888 and a root mean square error of 3.32 s. Analysis of the regression coefficients, corroborated by random-forest permutation importance, identified surface curvature and the choice of tool path as the principal determinants of trajectory efficiency, whilst tool-path choice and wound width governed scan time. Across the full dataset, 99.8 per cent of all trajectory waypoints were found to be kinematically feasible, and the mean singularity-avoidance score was 0.963.

For the class of wound-scanning problem here considered, regression models of modest parametric complexity are to be preferred over architectures of greater elaboration. The framework presented furnishes a reproducible computational foundation upon which experimental validation, real-time adaptive replanning, and three-dimensional wound reconstruction may be built.