Learning symbolic features for rule induction in computer aided diagnosis

In computer aided medical diagnosis (CAD), interpretability of learned models is an important concern. Unfortunately, the raw data used to train a model are often in sub-symbolic form (for instance, images), which makes the application of symbolic learning methods difficult. One way to alleviate this problem is to construct symbolic features that describe images, and learn to extract those features from raw images. The sub-symbolic part of the model is then limited to the lowest layer, making the model as a whole more interpretable. This paper presents a case study of how simple rule-based learners can be used to learn interpretable models from visual data by including a symbolic feature extraction step, in the domain of CAD. The symbolic representation is supported by literature and learned in the supervised way by means of deep learning. It turns out that the learned models are equally accurate as the black-box models that constitute the current state of the art.

Jaar van publicatie:2015

Toegankelijkheid:Open