Добавил:
Upload Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Medical Image Processing.pdf
Скачиваний:
26
Добавлен:
11.05.2015
Размер:
6.14 Mб
Скачать

182

L. Domanski et al.

Table 8.1 Speedup of linear feature detection algorithm after GPU acceleration. Results indicate how many times faster the NMS and gap filling ran on the GPU in isolation, as well as how many times faster the process ran overall

 

Speedup on GPU (times ×)

 

 

 

 

 

 

 

 

Img

NMS

Gap filling

Overall

 

1

13.3

4.8

3.0

 

2

13.7

8.3

3.3

 

3

13.6

8.6

3.3

 

4

9.0

8.6

3.3

 

5

7.9

3.5

2.7

 

6

8.0

6.9

3.0

 

8.3.5 Results for GPU Linear Feature Detection

We have also sped up the gap closing steps on the GPU, which involved parallelizing polar image generation and shortest path calculations. Various lower level code and memory optimisations were applied throughout. For example, loop unrolling has been applied to the NMS kernels [11], while texture fetches, a special kind of cached [12, 13]. GPU RAM access operation used in computer graphics, have been used to accelerate polar image generation. Discussion of these topics is beyond the scope of an introductory text. The reader may want to check [14] if his/her curiosity has been aroused.

Tests were performed on a Geforce GTX 260 GPU (NVIDA Corp., Santa Clara, CA, USA). The host computer consisted of a Xeon E5520 2.3GHz CPU (Intel Corp., Santa Clara, CA, USA) with 3GB of RAM running 32-bit Windows XP (Microsoft Corp., Redmond, WA, USA). Table 8.1 shows the performance of the linear feature detection algorithm after GPU accelerations for the images displayed in Fig. 8.7. The speedup for NMS is quite high, and is more significant for larger images (Fig. 8.7, img 1–3). The speedup for gap filling is also significant in the case of complex images (img 2–4, 6), but reduces for less complex images. These types of performance inconsistencies are common in parallel programs, where higher workloads help offset the costs of setup and communication that tend to take a constant amount of time or do not parallelize well.

Note that the overall speedup of the algorithm is much lower than the speedup of the accelerated steps in isolation due to the portion of the process consumed by the steps that were not accelerated. Since these steps consumed around 21–24% of the original CPU algorithm’s execution time for the test images, the maximum overall speedup we could hope for by accelerating NMS and gap filling is less than 4.7 times. The described GPU algorithms have achieved up to 70% of this theoretical speedup. Greater overall speedup will come from accelerating other parts of the algorithm.

8 High-Throughput Detection of Linear Features: Selected Applications...

183

8.4HCA-Vision, A Software Platform for Linear Feature Detection and Analysis

The algorithms detailed in Sects. 8.2 and 8.3 have been integrated in standalone software that rapidly and reproducibly quantifies linear features and their organization. It can be used both for linear feature detection and for more dedicated morphological analyses, such as neurite analysis. HCA-Vision was originally aimed for cell phenotyping and thus also provides functions for cell and nucleus detection and counting, cell scoring and sub-cellular analysis. It supports batch processing with a built-in database to store results, a batch result viewer and an ad hoc query builder for users to retrieve events of interest (e.g., identify all cells with neurite field above 50 μm2 and projecting in excess of 5 root neurites). Results from HCAVision have been shown to be more objective, reliable and reproducible than those from manual or semi-automated approaches [15].

8.4.1 Architecture and Implementation

HCA-Vision has been designed with three layers: the image processing layer written in C, a C++ wrapper layer, and the data and image presentation layer, written in C#. It also includes an optional web service for users to remotely view their batch processing results.

The GUI provides an assistant to guide the user in the process of choosing optimal parameters for the automated segmentation of cells and sub-cellular structures. It delivers quantitative measures and statistics that are biologically relevant, together with result images.

8.4.2 HCA-Vision Features

HCA-Vision enables a range of various activities including the detection of neurons, neurites, astrocytes, and other glial cells, even in the presence of varying background brightness, variable neurite staining, high cell densities, and high levels of neurite interconnections. It delivers quantitative measures and statistics that are biologically relevant, including measurements of neurite features at various levels of branching, neuron “webbiness,” and astrocyte “starriness” and roundness (see Table 8.3 for definitions).

To accommodate a wide range of applications, a step-by-step wizard guides the user through a process that fine-tunes individual parameter required for obtaining best results. At each step of the process, the user is given a semantic description of each parameter and controls their effect through user-friendly interfaces. As the processing has been optimized for speed, intermediate results are immediately presented on screen as each parameter is adjusted (Fig. 8.8).

184

L. Domanski et al.

Fig. 8.8 Results produced by HCA-Vision. (a) Original image. (b) Detected neurite segments. (c) Assignment of neurites to neuron bodies. (d) Labelled neurite branching structure

Once parameters are tuned, they can be saved into a profile. The user can load the saved profile to process individual images or batch process all images generated in an experiment. Summaries of batch processing results can be produced using the query facilities provided.

8.4.3 Linear Feature Detection and Analysis Results

HCA-vision detects linear features and reports their properties such as length, width, area and complexity. For example, with the neurite analysis module, neurites can be detected and their branching behaviour can be analysed, including the primary neurites in contact with the cell body and the number of layers of branching from these primary neurites into secondary, tertiary and quaternary branches (see Fig. 8.9).

Users can query and view the batch processing results using the batch result viewer (Fig. 8.10), including both the result images and measured features.

When HCA-Vision is used with microplates, the batch processing also generates a plate summary with normalized features for each well. The features extracted from the images captured from the same well are averaged to produce the well-based normalized statistics. A sample plate summary is shown in Table 8.2.

Соседние файлы в предмете [НЕСОРТИРОВАННОЕ]