Uncertainty-Driven Forest Predictors for Vertebra Localization and Segmentation

Abstract

Abstract. Accurate localization, identification and segmentation of vertebrae is an important task in medical as well as biological image analysis. The prevailing approach to solve such a task is to first generate pixel-independent features for each vertebra, e.g. via a random forest predictor, which are then fed into an MRF-based objective to infer the optimal MAP solution of a constellation model. We abandon this static, two-stage approach and mix feature generation with model-based inference in a new, more flexible, way. We evaluate our method on two data sets with different objectives. The first is semantic segmentation of a 21-part developing spine of zebrafish in microscopy images, and the second is localization and identification of vertebrae in benchmark human CT.

Mapping Auto-context Decision Forests to Deep ConvNets for Semantic Segmentation

“…The DF-initialized ConvNet achieved a Dice score of 0.66 after re-training, corresponding to a 10% relative improvement (see Figure 9(d) and Table 1(ConvNet)). This result matches previous State-of-the-Art results on this data set [39], but without the need for timeconsuming model-based inference. It's interesting to note that training a deep ConvNet with 8 hidden layers using hyperbolic tangent activation functions, and without batch normalization [21], is typically extremely difficult, but works well here, likely due to the good initialization of the network.…”

“…A common strategy in stacked classification is to introduce smoothing between the layers of the stack (e.g. [23,26,39]), and it appears that a similar strategy is naturally learned by the deep ConvNet.…”

Single-cell analysis of experience-dependent transcriptomic states in the mouse visual cortex

“…A labeled set of 23 images containing a total of ~2000 nuclei was used to train a stacked Random Forest (RF) classifier, using contextual "offset" features adapted from Refs. 49 and 50 , as well as newly developed circularity features. The RF was trained on three labels: background, nuclei contour, and nuclei, therefore producing three respective probability maps.…”

Bots for Software-Assisted Analysis of Image-Based Transcriptomics

“…For example, in the Auto-context model, new features are generated by sampling predictions over a contextual grid, and as a result the stacked classifier is able to learn stereotypical class layouts [26]. This approach has been applied to numerous different tasks, including facade segmentation [11], and bio-medical image segmentation [13,18]. Here we adapt this idea to address the challenge of splitting touching nuclei by introducing "circularity features" that leverage prior knowledge of nucleus shape to better identify their boundaries.…”