Recognition and morphometry Models 3.0 (2000-2001)

Recognition and morphometry Models 3.0 (2011)

Models evolution:

The 2011 ANN models are a just mode modern variant of the older 2000-2001 models, which differ on the following points:

  • They are using SVR models (regression SVM, support vector machines) instead of MLP (multi-layer perceptrons), which provide sensibly the same performance but are easier to handle in our machine learning framework.
  • They have been trained on the 2008 learning database of 62 subjects, which means that they are more accurate, and follow the new sulci nomenclature, like the SPAM-based recognition models.

    • In other respects, they are similar to the 2000-2001 models.

    Performance

    The current models perform about 5% better than the older ones. Here are the error rates estimated via a Leave-One-Out cross-validation:

    Side: SI error: Mass error:
    left 17.11 ± 4.31 19.26 ± 4.53
    right 16.86 ± 4.78 19.02 ± 4.89

    Short description of the descriptors for sulci:

    The descriptors are the same as the older models.