This article illustrates the process of performing reproducible research with existing open source tools.

Reproducible Research Training at Scipy 2014

This document describes a module that extends the Insight Toolkit, ITK, which reads DICOM Spatial Rogistration Object files in itk::Transform's. Currently, DICOM files are read by applying the DCMTK library as a backend. An itk::DCMTKTransformIO class can be registration with the IO factory mechanism so itk::TransformFileReadertemplate will recognize and read these files.

This paper is accompanied with the source code, input data, parameters and output data that the authors used for validating the algorithm described in this paper. This adhere to the fundamental principle that scientific publications must facilitate reproducibility of the reported results.

DICOM Spatial Transform IO in the Insight Toolkit

This paper describes how to compile and use the Insight Toolkit Kinetic Analysis library (itk::ka), to perform analysis of dynamic medical images, along with a brief overview of the library source code. Supplied, are a set of interfaces to facilitate application development, application implementing a standard kinetic analysis pipeline (koala), and some sample data. Only versions of ITK4 and higher are currently supported.

InsightToolkit Kinetic Analysis (itk::ka) Library

The segmentation of parotid glands in CT scans of patients with head and neck cancer is an essential part of treatment planning. We introduce a new method for the automatic segmentation of parotid glands that extends existing patch-based approaches in three ways: (1) we promote the use of image features in combination with patch intensity values to increase discrimination; (2) we work with larger search windows than established methods by using an approximate nearest neighbor search; and (3) we demonstrate that location information is a crucial discriminator and add it explicitly to the description. In our experiments, we compare a large number of features and introduce a new multi-scale descriptor. The best performance is achieved with entropy image features in combination with patches and location information.

On the Importance of Location and Features for the Patch-Based Segmentation of Parotid Glands

Real-time 3D echocardiography has already been shown to be an accurate tool for left ventricular (LV) volume assessment. However, LV border identification remains a challenging task, mainly because of the low contrast of the images combined with drop-out artifacts. Many (semi-)automatic algorithms have been proposed to segment the LV border, but a systematic and fair comparison between such methods has so far been impossible due to a lack of publicly available common database. The aim of this MICCAI challenge was to gather researchers around the field of LV segmentation in 3D cardiac ultrasound by providing a common database to compare algorithms directly. The proposed platform will allow a consistent evaluation and ranking of the current state-of-the-art segmentation solutions and will contribute to a faster clinical translation of groundbreaking technical advances. The purpose of this paper is to describe the technical aspects of the generation of the database, give an overview of the ranking strategy and the outline of the challenge itself.

Challenge on Endocardial Three-dimensional Ultrasound Segmentation (CETUS)

Within the framework of the CETUS challenge, forty-five 3D echocardiographic datasets have been acquired and segmented independently by three clinical experts from different hospitals. The goal was to generate a well-established ground truth of validated expert contours on this broad range of images from different ultrasound vendors, for a number of common pathologies. Image data were acquired and segmented according to a specifically designed protocol. Since there is no clear standard or guideline for segmentation for 3DUS, we defined a tracing consensus which results in clinically acceptable and reproducible contours. Tracing was performed in four longitudinal and five transversal 3D-derived 2D planes in ED and ES. 3D contours were constructed from these tracings. If the contours or their clinical parameters differed by more than a predefined level, the tracings were compared and the experts
would reach a consensus interpretation on the best segmentation. One or more experts would then adapt their tracings. Final distance differences in contours were 0.77±0.18mm for the training set and 0.82±0.27mmfor the testing set. For the training set, 69% of contours were retraced. For the testing set, 76% of contours were retraced. The used protocol resulted in well-established ground truth contours.

Clinical Expert Delineation of 3D Left Ventricular Echocardiograms for the CETUS Segmentation Challenge

We present a novel method for segmentation and tracking of the left ventricle (LV) in 4D ultrasound sequences using a combination of automatic segmentation at the end-diastolic frame and tracking using both a global optical flow-based tracker and local block matching. The core novelty of the proposed algorithm relies on the recursive formulation of the block-matching problem, which introduces temporal consistency on the patterns being tracked. The proposed method offers a competitive solution, with average segmentation errors of 2.29 and 2.26mm in the training (#=15) and testing (#=15) datasets respectively.

Fast Tracking of the Left Ventricle Using Global Anatomical Affine Optical Flow and Local Recursive Block Matching

In this paper, we present a fast and interactive graph cut method for 3D segmentation of the endocardial wall of the left ventricle (LV) given 3D echocardiographic images. This is a challenging task due to the poor contrast and the low signal-to-noise ratio typical of echocardiographic images. The method is carried out in 3 steps. First, 3D sampling of the LV cavity is made in a spherical-cylindrical coordinate system. Then, a gradient-based energy term is assigned to each voxel, some of which being given an infinite energy to make sure the resulting volume passes through key anatomical points. Then, a graph-cut procedure provides delineation of the endocardial surface. Results obtained on 30 exams from the 2014 CETUS MICCAI challenge dataset reveal that our method takes between 5 and 10 seconds to segment a 3D volume with an overall mean surface distance lower than 2.3 mm and an ejection fraction error of less than 5% compared to a manual tracing by an expert.

Automatized Evaluation of the Left Ventricular Ejection Fraction from Echocardiographic Images Using Graph Cut

Segmentation of the left ventricle endocardium in 3D echocardiography is a critical step for the diagnosis of heart disease. Although recent work has shown effective endocardial edge detection, these techniques still preserve spurious anatomical edge responses that undermine overall ventricle segmentation. In this paper we propose a robust semiautomatic framework based on 2D structured learning that facilitates full 3D model-based endocardial segmentation. This method is evaluated on 30 publicly available datasets from different brands of ultrasound machines. Results show that the proposed method accurately finds the endocardium and effectively converges an explicit and continuous surface model to it.

Endocardial Segmentation using Structured Random Forests in 3D Echocardiography

In this paper, we present the use of a generic image segmentation method, namely a succession of Random Forest classifiers in an autocontext framework, for the MICCAI 2014 Challenge on Endocardial 3D Ultrasound Segmentation (CETUS). The proposed method segments each frame independently in 90 sec, without requiring temporal information such as end-diastolic or end-systolic time points nor any registration. For better segmentation accuracy, non-local means denoising can be applied to the images at the cost of an increased run-time. The mean Dice score on the testing dataset was 84.4% without denoising and 86.4% with denoising. The originality of our approach lies in the introduction of two classes, the myocardium and the mitral valve, in addition to the left ventricle and the background classes, in order to gain contextual information for the segmentation task.