Multi-organ Localization Combining Global-to-Local Regression and Confidence Maps

Abstract

We propose a method for fast, accurate and robust localization of several organs in medical images. We generalize global-to-local cascades of regression forests [1] to multiple organs. A first regressor encodes global relationships between organs. Subsequent regressors refine the localization of each organ locally and independently for improved accuracy. We introduce confidence maps, which incorporate information about both the regression vote distribution and the organ shape through probabilistic atlases. They are used within the cascade itself, to better select the test voxels for the second set of regressors, and to provide richer information than the classical bounding boxes thanks to the shape prior. We demonstrate the robustness and accuracy of our approach through a quantitative evaluation on a large database of 130 CT volumes.

Scale-Adaptive Forest Training via an Efficient Feature Sampling Scheme

“…While they can be specifically designed for an application if some domain knowledge is available, a popular and effective approach [3][4][5][6][7][8][9] consists in extracting a large number of low-level Haar-like features corresponding to visual cues at offset locations. Each Haar-like feature is characterized by a parameter vector λ ∈ Λ which defines, for each pixel p, a certain type of contextual information x λ (p) ∈ R as follows.…”

“…Recent applications of the forest framework to the medical field include multi-organ segmentation within computed tomography (CT) volumes [3], segmentation of the midbrain in transcranial ultrasound volumes [4], multi-organ localization in magnetic resonance (MR) [5] and CT [6] data, semantic labeling of brain structures in MR scans [7], depth video classification to quantify the progression of multiple sclerosis [8], and localization of anatomical landmarks within hand MR scans [9].…”

“…For this reason, incorporating multi-scale information during training is of great interest, and several approaches have been proposed to achieve this objective. A common but computationally costly strategy is to perform several independent learning stages at various scales and combine their outputs at prediction time [6,9,10]. Geremia et al [11] explicitly create a hierarchy of supervoxels and refine the representation when necessary during the forest training.…”

“…Our approach builds on the generic Haar-like features [12] which demonstrated lately high performance for a variety of medical objectives and imaging modalities [3][4][5][6][7][8][9]. Instead of sampling Haar-like features uniformly at each node, we sample them sequentially in a fine-to-coarse fashion.…”

Light Random Regression Forests for automatic multi-organ localization in CT images

“…Since its introduction, RRFs have been successfully used as the first localization step in fully automatic segmentation of right and left kidneys [5], liver [6], multiple organs (liver, right kidney, left kidney, spleen, gallbladder, and stomach) [7], hip joint [8], and detection, grading and classification of coronary stenoses [9] in CT volumes.…”