Stable and discriminating features are predictive of cancer presence and Gleason grade in radical prostatectomy specimens: a multi-site study

Abstract

Site variation in fixation, staining, and scanning can confound automated tissue based image classifiers for disease characterization. In this study we incorporated stability into four feature selection methods for identifying the most robust and discriminating features for two prostate histopathology classification tasks. We evaluated 242 morphology features from N = 212 prostatectomy specimens from four sites for automated cancer detection and grading. We quantified instability as the rate of significant cross-site feature differences. We mapped feature stability and discriminability using 188 non-cancerous and 210 cancerous regions via 3-fold cross validation, then held one site out, creating independent training and testing sets. In training, one feature set was selected only for discriminability, another for discriminability and stability. We trained a classifier with each feature set, testing on the hold out site. Experiments were repeated with 117 Gleason grade 3 and 112 grade 4 regions. Stability was calculated across non-cancerous regions. Gland shape features yielded the best stability and area under the receiver operating curve (AUC) trade-off while co-occurrence texture features were generally unstable. Our stability-informed method produced a cancer detection AUC of 0.98 ± 0.05 and increased average Gleason grading AUC by 4.38%. Color normalization of the images tended to exacerbate feature instability.

First-Stage Prostate Cancer Identification on Histopathological Images: Hand-Driven versus Automatic Learning

“…The first step, before addressing the feature extraction stage, is to obtain separately the four main components that appear in the H&E prostate images, i.e., lumens, nuclei, cytoplasm and stroma, in order to compute different features from each component, as well as from the relation between them. Similarly to [ 15 , 16 ], we apply clustering algorithms based on the k -means technique to carry out the identification of each tissue component. However, unlike the previous studies that only used the RGB image as input for the clustering step, we use different colour spaces depending on the component mask that we want to extract.…”

“…In this work, we make use of two kinds of descriptors for encoding the textural information related to the artefacts, benign and pathological glands. On the one hand, we use Gray-Level Co-occurrence Matrix (GLCM), similarly to Leo et al [ 15 ], who calculated a co-occurrence matrix to obtain information about the glands’ orientation. In other studies, such as [ 11 , 13 , 32 ], the authors also applied GLCM-based techniques but on histological regions, instead of gland units.…”

“…Machine-learning classifiers . In order to address the classification stage through hand-driven learning methods, we make use of two different classifiers widely utilised in the literature [ 6 , 9 , 10 , 15 ]. Specifically, we optimise a Support Vector Machine (SVM) classifier, which allows for finding the optimal hyperplane h that separates two regions of the input space maximising the distance between two support vectors , one of each class [ 38 ].…”

“…From the segmented regions of lumens and nuclei, the authors applied a total of 670 morphological features to detect prostate cancer and they reported an Area Under the ROC Curve (AUC) of 0.98 using a multi-view boosting classifier. In another recent study [ 15 ], the researchers made use of the Nuclei-Lumen Association (NLA) algorithm, proposed in [ 16 ], to carry out the gland unit segmentation. Then, the authors extracted features based on graphs and texture, as well as shape and orientation disorder of the glands, with the aim of distinguishing between cancerous and non-cancerous regions.…”

“…Regarding the approach related to the hand-driven learning, we incorporate in a novel way a hand-crafted feature extraction stage based on four global kinds of descriptors. We not only use morphological and textural descriptors as the majority of studies [ 4 , 8 , 11 , 13 , 15 ], but also descriptors related to the fractal dimension, like [ 5 , 12 , 22 ], and contextual features, similarly to [ 6 , 16 ]. Until now, each kind of descriptor had been implemented separately, but, in this paper, we build a hybrid feature vector able to encode all the relevant information included in the gland units.…”