A Bottom-Up Approach for Pancreas Segmentation Using Cascaded Superpixels and (Deep) Image Patch Labeling

Abstract

Abstract-Robust automated organ segmentation is a prerequisite for computer-aided diagnosis (CAD), quantitative imaging analysis, detection of pathologies and surgical assistance. For anatomical high-variability organs such as the pancreas, previous segmentation approaches report low accuracies in comparison to well studied organs like the liver or heart. We present a fully-automated bottom-up approach for pancreas segmentation in abdominal computed tomography (CT) scans. The method is based on a hierarchical cascade of information propagation by classifying image patches at different resolutions and cascading (segments) superpixels. There are four stages in the system: 1) decomposing CT slice images as a set of disjoint boundarypreserving superpixels; 2) computing pancreas class probability maps via dense patch labeling; 3) classifying superpixels by pooling both intensity and probability features to form empirical statistics in cascaded random forest frameworks; and 4) simple connectivity based post-processing. The dense image patch labeling is conducted by two schemes: efficient random forest classifier on image histogram, location and texture features; and more expensive (but with better specificity) deep convolutional neural network classification, on larger image windows (i.e., with more spatial contexts). Over-segmented 2D CT slices by the Simple Linear Iterative Clustering approach are adopted through model/parameter calibration and labeled at the superpxiel level for positive (pancreas) or negative (non-pancreas background) classes. Evaluation of the approach is done on a database of 80 manually segmented CT volumes in six-fold cross-validation. Our achieved results are comparable, or better than the state-of-theart methods (evaluated by "leave-one-patient-out"), with a Dice coefficient of 70.7% and Jaccard Index of 57.9%. In addition, the computational efficiency has been drastically improved in the order of 6 ∼ 8 minutes, comparing with others of ≥ 10 hours per testing case. The segmentation framework using deep patch labeling confidences is also more numerically stable, reflected by the smaller performance metric standard deviations. Finally, we implement a multi-atlas label fusion (MALF) approach for pancreas segmentation using the same dataset. Under six-fold cross-validation, our bottom-up segmentation method significantly outperforms its MALF counterpart: 70.7 ± 13.0% versus 52.51 ± 20.84% in Dice coefficients.

Repeatability and Reproducibility of Pancreas Volume Measurements Using MRI

“…One intriguing alternative to the use of multiple readers is the use of machine learning techniques for computer-driven pancreas segmentation 16,17 . Automatic segmentation circumvents inter-reader variability, although we have found this variation to be modest.…”

CT image segmentation of bone for medical additive manufacturing using a convolutional neural network

“…This has motivated many researchers to use CNNs for various medical segmentation tasks, such as the segmentation of brain tissue [24][25][26], prostate [27], bone [28,29], and tumors [30][31][32][33] in MR images. Furthermore, multiple studies have been conducted on the segmentation of kidneys [34] and the pancreas [35][36][37] in CT scans. A few studies have investigated the use of CNNs for bone segmentation in CT scans.…”

Deep Learning in Medical Image Analysis

“…However, networks with vectorized inputs were also successfully applied to different medical applications (47, 25, 27, 29, 48, 24, 49, 50). Along with deep neural networks, deep generative models (51) such as deep belief networks and deep Boltzmann machines that are the probabilistic graphical models with multiple layers of hidden variables have also been successfully applied to brain disease diagnosis (43, 25, 52, 29), lesion segmentation (53, 45, 32, 54), cell segmentation (33, 55, 34, 56), image parsing (57, 58, 59), and tissue classification (31, 46, 22, 44). …”