Randomized Tree Ensembles for Object Detection in Computational Pathology

Abstract

Randomized tree ensembles for object detection in computational pathology AbstractModern pathology broadly searches for biomarkers which are predictive for the survival of patients or the progression of cancer. Due to the lack of robust analysis algorithms this work is still performed manually by estimating staining on whole slides or tissue microarrays (TMA). Therefore, the design of decision support systems which can automate cancer diagnosis as well as objectify it pose a highly challenging problem for the medical imaging community. In this paper we propose Relational Detection Forests (RDF) as a novel object detection algorithm, which can be applied in an off-the-shelf manner to a large variety of tasks. The contributions of this work are twofold: (i) we describe a feature set which is able to capture shape information as well as local context. Furthermore, the feature set is guaranteed to be generally applicable due to its high flexibility.(ii) we present an ensemble learning algorithm based on randomized trees, which can cope with exceptionally high dimensional feature spaces in an efficient manner. Contrary to classical approaches, subspaces are not split based on thresholds but by learning relations between features. The algorithm is validated on tissue from 133 human clear cell renal cell carcinoma patients (ccRCC) and on murine liver samples of eight mice. On both species RDFs compared favorably to state of the art methods and approaches the detection accuracy of trained pathologists. Abstract. Modern pathology broadly searches for biomarkers which are predictive for the survival of patients or the progression of cancer. Due to the lack of robust analysis algorithms this work is still performed manually by estimating staining on whole slides or tissue microarrays (TMA). Therefore, the design of decision support systems which can automate cancer diagnosis as well as objectify it pose a highly challenging problem for the medical imaging community. In this paper we propose Relational Detection Forests (RDF) as a novel object detection algorithm, which can be applied in an off-the-shelf manner to a large variety of tasks. The contributions of this work are twofold: (i) we describe a feature set which is able to capture shape information as well as local context. Furthermore, the feature set is guaranteed to be generally applicable due to its high flexibility. (ii) we present an ensemble learning algorithm based on randomized trees, which can cope with exceptionally high dimensional feature spaces in an efficient manner. Contrary to classical approaches, subspaces are not split based on thresholds but by learning relations between features. The algorithm is validated on tissue from 133 human clear cell renal cell carcinoma patients (ccRCC) and on murine liver samples of eight mice. On both species RDFs compared favorably to state of the art methods and approaches the detection accuracy of trained pathologists. Randomized Tree Ensembles for Object Detection in Computational Pathology

Computational pathology: Challenges and promises for tissue analysis

“…It is an ongoing project in kidney cancer research conducted at the University Hospital Zurich and ETH Zurich. Parts of it were published in [100] and [42], where also algorithmic details of the computational approach can be found. Figure 8 depicts a schematic overview of the project subdivided into the three main parts which are discussed in the following:…”

“…Relational Detection Forests [42] provide one possibility to overcome this problem of information loss. Especially designed for detection of cell nuclei in histological slides, they are based on the concept of randomized trees [43].…”

“…The feature base proposed in [42] is defined as follows: The coordinates of two rectangles R 1 and R 2 are sampled uniformly within a predefined window size w:…”

“…Approaches like AdaBoost [46] which yield very good results for up to hundreds of thousands of features are not applicable any more. These problems can be overcome by employing randomized algorithms [42,47] where features are sampled randomly for learning classifiers on these random projections.…”

Generalizing the Majority Voting Scheme to Spatially Constrained Voting

“…As corresponding examples, we can mention neural networks [2], [3], decision trees [4], sets of rules [5] and other models [6], [7], [8]. As a specific application field, now we will focus on object detection in digital images which is a vivid field [9], [10], [11], as well.…”