possible to make up for the gap in the doctors’ diagnostic level and improve the
detection of early symptoms of pulmonary nodules in a wider area.
CAD system is generally composed of five main stages: acquisition of data, preprocessing, lung segmentation, nodule detection and the reduction of false positives [2].
The first stage is responsible for obtaining the images as the input of CAD system. The
preprocessing stage includes a series of image processing to improve the quality of the
image or to compress the amount of input data to improve the result of later stage. The
lung segmentation stage separates the learning area from other organs and tissues in the
ray image to reduce the computational cost of the next stage. The nodule detection
stage determines the location of the lung nodules in the segmented image as candidates.
The reduction of false positives stage use machine-learning algorithms to eliminate
false nodules from the candidate nodules.
In this work, our purpose is to parallelize the stages before reduction of false
positives, which can accelerate the calculation speed without affecting the output
results. To achieve this purpose, we developed three concepts to deal with the existing
system [3].
1. Data concept. Because the 3D lung CT image is composed of number of 2D image
slices, we treat each 2D image slice as one processing unit. We group the 2D image
slices and process these groups at the same time.
2. Instruction concept. Preprocessing, lung segmentation and nodule detection are
mutually independent, so we spread these stages across different execution units and
perform parallel processing on multiple data in a pipelined manner.
3. Data transmission concept. We use the bus structure to optimize data transmission
and simplify the system structure.
2 Parallelization Scheme Design
When designing the parallelization scheme, we considered the differences in various
scenarios. For medical institutions with a complete data center and powerful computing
capabilities, the CAD systems can be deployed in a local environment. In this case, we
designed a parallelization scheme for the local environment. However, for medical
institutions without the capabilities, when they need to conduct remote consultation,
they have to transmit data to the remote computer cluster. In this case, we consider
about how to transmit data efficiently and reliably in computer cluster. We designed
three schemes to meet them: static grouping-based scheme, pipeline-based scheme and
bus-based parallelization scheme.
2.1 Static Grouping-Based Parallelization Scheme
The static grouping-based scheme refers to the fixed division of slices into the same
number of group as the execution unit, and each execution unit only needs to care about
the grouped slices assigned to it. There is no data transmission between the execution
units, they only need to receive input and send output to control unit. For the control
unit, its main job is to group the 2D image slices to be processed. In order to ensure that
Parallelization of Preprocessing for Automatic Detection …
585
detection of early symptoms of pulmonary nodules in a wider area.
CAD system is generally composed of five main stages: acquisition of data, preprocessing, lung segmentation, nodule detection and the reduction of false positives [2].
The first stage is responsible for obtaining the images as the input of CAD system. The
preprocessing stage includes a series of image processing to improve the quality of the
image or to compress the amount of input data to improve the result of later stage. The
lung segmentation stage separates the learning area from other organs and tissues in the
ray image to reduce the computational cost of the next stage. The nodule detection
stage determines the location of the lung nodules in the segmented image as candidates.
The reduction of false positives stage use machine-learning algorithms to eliminate
false nodules from the candidate nodules.
In this work, our purpose is to parallelize the stages before reduction of false
positives, which can accelerate the calculation speed without affecting the output
results. To achieve this purpose, we developed three concepts to deal with the existing
system [3].
1. Data concept. Because the 3D lung CT image is composed of number of 2D image
slices, we treat each 2D image slice as one processing unit. We group the 2D image
slices and process these groups at the same time.
2. Instruction concept. Preprocessing, lung segmentation and nodule detection are
mutually independent, so we spread these stages across different execution units and
perform parallel processing on multiple data in a pipelined manner.
3. Data transmission concept. We use the bus structure to optimize data transmission
and simplify the system structure.
2 Parallelization Scheme Design
When designing the parallelization scheme, we considered the differences in various
scenarios. For medical institutions with a complete data center and powerful computing
capabilities, the CAD systems can be deployed in a local environment. In this case, we
designed a parallelization scheme for the local environment. However, for medical
institutions without the capabilities, when they need to conduct remote consultation,
they have to transmit data to the remote computer cluster. In this case, we consider
about how to transmit data efficiently and reliably in computer cluster. We designed
three schemes to meet them: static grouping-based scheme, pipeline-based scheme and
bus-based parallelization scheme.
2.1 Static Grouping-Based Parallelization Scheme
The static grouping-based scheme refers to the fixed division of slices into the same
number of group as the execution unit, and each execution unit only needs to care about
the grouped slices assigned to it. There is no data transmission between the execution
units, they only need to receive input and send output to control unit. For the control
unit, its main job is to group the 2D image slices to be processed. In order to ensure that
Parallelization of Preprocessing for Automatic Detection …
585