Weakly Supervised Group-Wise Model Learning Based on Discrete Optimization

Abstract

Abstract. In this paper we propose a method for the weakly supervised learning of sparse appearance models from medical image data based on Markov random fields (MRF). The models are learnt from a single annotated example and additional training samples without annotations. The approach formulates the model learning as solving a set of MRFs. Both the model training and the resulting model are able to cope with complex and repetitive structures. The weakly supervised model learning yields sparse MRF appearance models that perform equally well as those trained with manual annotations, thereby eliminating the need for tedious manual training supervision. Evaluation results are reported for hand radiographs and cardiac MRI slices.

Constructing part-based models for groupwise registration

“…We achieve an accuracy of 0.7mm on the positioning of the chosen parts. This is significantly better than results quoted by Donner et al [13] (approx. 1.5mm, though on a different, smaller dataset).…”

“…For instance, Donner et al [12] demonstrated how a sophisticated parts + geometry model can accurately locate points in such images. In further work [13] they showed that such a model can be constructed automatically from a set of images in which only one is manually annotated. However, the method was only evaluated on a small set of 12 hand radiographs.…”

One-shot learning of anatomical structure localization models

“…This yields results which represent parts of the image in more detail than others, and the location of anatomical landmarks has to be inferred from the landmarks chosen by the approach through interpolation. [8] employ incremental model building similar to our work, but again the underlying method is restricted to landmarks obtained through standard interest point detectors, such as Harris corners, instead of anatomical landmarks. Such interest points pose a delicate sensitivity / number of landmarks trade-off, resulting in inaccurate or very slow model matching / hypothesis generation, especially on 3D data.…”

Learning New Parts for Landmark Localization in Whole-Body CT Scans

“…However, that approach represents shape correspondences using only sparse interest points and a global active shape model, which cannot deal efficiently with multiple candidates and is more easily affected by local perturbations. Donner et al [8] learn a GM of shape and appearance, which relies on a single manually annotated training image to define the structure of interest, and iteratively include high confidence matches in the emerging model. In subsequent work [7], the authors introduced a method that is similar to our baseline model in that it does not rely on predefined interest point detector.…”