A computer-aided detection (CAD) system for the identification of pulmonary nodules in low-dose multi-detector helical Computed Tomography (CT) images was developed in the framework of the MAGIC-5 Italian project. One of the main goals of this project is to build a distributed database of lung CT scans in order to enable automated image analysis through a data and cpu GRID infrastructure. The basic modules of our lung-CAD system, a dot-enhancement filter for nodule candidate selection and a neural classifier for false-positive finding reduction, are described. The system was designed and tested for both internal and sub-pleural nodules. The results obtained on the collected database of low-dose thin-slice CT scans are shown in terms of free response receiver operating characteristic (FROC) curves and discussed.
Deep Dive into An automated system for lung nodule detection in low-dose computed tomography.
A computer-aided detection (CAD) system for the identification of pulmonary nodules in low-dose multi-detector helical Computed Tomography (CT) images was developed in the framework of the MAGIC-5 Italian project. One of the main goals of this project is to build a distributed database of lung CT scans in order to enable automated image analysis through a data and cpu GRID infrastructure. The basic modules of our lung-CAD system, a dot-enhancement filter for nodule candidate selection and a neural classifier for false-positive finding reduction, are described. The system was designed and tested for both internal and sub-pleural nodules. The results obtained on the collected database of low-dose thin-slice CT scans are shown in terms of free response receiver operating characteristic (FROC) curves and discussed.
Lung cancer is one of the most relevant public health issues. Despite significant research efforts and advances in the understanding of tumour biology, there was no reduction of the mortality over the last decades.
Lung cancer most commonly manifests itself with the formation of non-calcified pulmonary nodules. Computed Tomography (CT) is the best imaging modality for the detection of small pulmonary nodules, particularly since the introduction of the helical technology 1 . However, the amount of data that need to be interpreted in CT examinations can be very large, especially when multi-detector helical CT and thin collimation are used, thus generating up to about 300 two-dimensional images per scan, corresponding to about 150 MB. In order to support radiologists in the identification of early-stage pathological objects, researchers have recently begun to explore computer-aided detection (CAD) methods in this area.
Among the approaches that are being tried to reduce the mortality of lung cancer is the implementation of screening programs for the subsample of the population with higher risk of developing the disease. The First Italian Randomized Controlled Trial (ITALUNG-CT) that aims to study the potential impact of screening on a high-risk population using low-dose helical CT was recently started 2 .
A CAD system for small pulmonary nodule identification, based on the analysis of images acquired from the Pisa centre of the ITALUNG-CT trial, was developed in the framework of the MAGIC-5 collaboration funded by Istituto Nazionale di Fisica Nucleare (INFN) and Ministero dell’Università e della Ricerca (MIUR). The system is based on a dotenhancement filter for the identification of nodule candidates and a neural network based classification module for the reduction of the number of false-positive (FP) findings per scan. *ilariagori@gmail.com; phone +39 340 6220624
A low-dose lung CT dataset was acquired from the Pisa centre of the ITALUNG-CT trial, the First Italian Randomized Controlled Trial for the screening of lung cancer 2 . The CT scans are acquired with a 4 slices spiral CT scanner according to a low-dose protocol (screening setting: 140 kV, 20 mA), with a 1.25 mm slice collimation.
The dataset used for this study consists of 39 low-dose CT scans. Each scan is a sequence of slices stored in DICOM (Digital Imaging and COmmunications in Medicine) format. The average number of slices per scan is about 300 with 512×512 pixel matrix, a pixel size ranging from 0.53 to 0.74 mm and 12 bit grey levels. The scans were annotated by experienced radiologists, by using a visualization and annotation tool we developed, described in the following subsection.
Non-calcified solid nodules only, with a diameter greater than 5 mm, were considered in this study, whereas groundglass opacities were excluded. The dataset consists of 102 annotated nodules, 75 internal nodules belonging to 34 out of the 39 available scans and 27 sub-pleural nodules belonging to 20 out of the 39 available scans.
Examples of internal and sub-pleural nodules extracted from our screening database are shown in Fig. 1.
In order to be able to insert into our database the annotations required to precisely identify the position of every diagnosed nodule, we had to develop an annotation tool, i.e. a software application with a graphical user interface (GUI) where the radiologist can explore the CT data, identify eventual nodules and annotate them using a standard classification scheme. This tool is built upon the MG framework, and thus strongly integrated with the ROOT 3 platform, allowing access to GRID-enabled distributed resources. The tool could possibly evolve into the main interface for the complete lung CAD system.
The annotation tool can read three-dimensional CT data scans stored in DICOM format, and will read and write annotation lists in a simple text-based format. At the moment, since this tool was developed in order to validate the results of the lung CAD algorithms, an annotation is simply associated to a spherical region of the dataset, identified by the position of the center and its radius. This is all the information currently required by the CAD validation tools. If different requirements shall arise, the modular organization of the software will allow other selection methods to be implemented (manual or computer-assisted boundary identification, for example).
The graphical interface allows the radiologist to visualize axial slices of the CT data (the standard visualization mode that radiologists are used to) (see Fig. 2). Standard imaging and navigational controls are provided, such as zoom/pan functionality and dynamic range control (the so-called window selection). Once the radiologist identifies a suspect nodule, they can select that region using a click-and-drag interface: a pop-up window allows classification of the nodule according to a standard set of parameters (morphological, and eventually clinical). The selected re
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
