• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site

Article

Pattern recognition in low-resolution instrumental tactile imaging

Nersisyan S., Rakhmatulin Y.

Background. Tactile perception is an essential source of information. However instrumental registration and automated analysis of tactile data is still at an initial point of the development. Recently a Medical Tactile Endosurgical Complex (MTEC) has been introduced into clinical practice as a universal instrument for intrasurgical registration of tactile images. Images registered by MTEC have very limited resolution both in terms of a number of tactile pixels and a number of discretization levels. In this study we investigated whether this resolution is sufficient for reliable pattern recognition. Methods. Our study used a set of artificial samples which included six sample types. In particular, four of these types directly tested the ability to discriminate patterns with the same embedment projection sizes but different curvatures, or similar curvatures but different projection sizes. Two widely used machine learning methods were evaluated: random forests and k-nearest neighbors. These methods were applied to points representing registered tactile images in a relatively low-dimensional feature space. Additionally an in-silico cloning of images was used to increase classification reliability. Results. Both classification methods – random forests and k-nearest neighbors – showed good classification reliability with accuracy 68.6% and 72.9% on the validation set, respectively. These values are more than four times higher than an accuracy of six-class “random classifier”. Random forests additionally provided evaluation of importance of features used for classification. Conclusion. Despite poor resolution of tactile images registered by MTEC a combination of conventional machine learning methods with a specific feature set and specific tricks provides highly reliable results of automated analysis of these images even in case of nontrivial tasks such as sample classification with very similar classes.