Posted on

auto ching?n seeds

Auto ching?n seeds
In many medical image segmentation applications, one of the major drawbacks for most of the algorithms is that they require deliberate initialization, which becomes impractical when dealing with large numbers of images, this issue has been addressed in some of the existing works already. In Lu’s work [32, 33], a single click was required for live wire algorithm and Yan’s work [44] also shows it can find the minimal path for an object from a single starting point. One of our contributions in the paper is that the SCES requires only one initialization (one starting point).

Automated Delineation of Lung Tumors from CT Images Using a Single Click Ensemble Segmentation Approach

Yuhua Gu

1 Department of Imaging, H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida 33612. USA

Virendra Kumar

1 Department of Imaging, H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida 33612. USA

Lawrence O Hall

2 Department of Computer Science and Engineering, University of South Florida, Tampa, Florida 33620. USA

Dmitry B Goldgof

2 Department of Computer Science and Engineering, University of South Florida, Tampa, Florida 33620. USA

Ching-Yen Li

2 Department of Computer Science and Engineering, University of South Florida, Tampa, Florida 33620. USA

René Korn

3 Definiens AG, Trappentreustraße 1, 80339 München, Germany

Claus Bendtsen

4 DECS, AstraZeneca, 50S27 Mereside, Alderley Park, Macclesfield, Cheshire SK10 4TG, UK

Emmanuel Rios Velazquez

5 Departments of Radiation Oncology, University Hospital Maastricht, Maastricht, Netherlands

Andre Dekker

5 Departments of Radiation Oncology, University Hospital Maastricht, Maastricht, Netherlands

Hugo Aerts

5 Departments of Radiation Oncology, University Hospital Maastricht, Maastricht, Netherlands

Philippe Lambin

5 Departments of Radiation Oncology, University Hospital Maastricht, Maastricht, Netherlands

Xiuli Li

6 Medical Image Processing Group, State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China

Jie Tian

6 Medical Image Processing Group, State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China

Robert A Gatenby

1 Department of Imaging, H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida 33612. USA

Robert J Gillies

1 Department of Imaging, H. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida 33612. USA

Abstract

A single click ensemble segmentation (SCES) approach based on an existing “Click&Grow” algorithm is presented. The SCES approach requires only one operator selected seed point as compared with multiple operator inputs, which are typically needed. This facilitates processing large numbers of cases. Evaluation on a set of 129 CT lung tumor images using a similarity index (SI) was done. The average SI is above 93% using 20 different start seeds, showing stability. The average SI for 2 different readers was 79.53%. We then compared the SCES algorithm with the two readers, the level set algorithm and the skeleton graph cut algorithm obtaining an average SI of 78.29%, 77.72%, 63.77% and 63.76% respectively. We can conclude that the newly developed automatic lung lesion segmentation algorithm is stable, accurate and automated.

1. Introduction

Lung cancer has become one of the most significant diseases in human history. The World Health Organization estimates the worldwide death toll from lung cancer will be 10,000,000 by 2030. The 5-year survival rate for advanced Non Small Cell Lung Cancer (NSCLC) [1] remains disappointingly low. It has been hypothesized that quantitative image feature analysis can improve diagnostic/prognostic or predictive accuracy, and therefore will have an impact on a significant number of patients [2]. In the current study, standard-of-care clinical computed tomography (CT) scans were used for image feature extraction. In order to reduce variability for feature extraction, the first and essential step is to accurately delineate the lung tumors. Accurate delineation of lung tumors is also crucial for optimal radiation oncology. A common approach to delineate tumor from CT scans involves radiologists or radiation oncologists manually drawing the boundary of the tumor. In the majority of cases, manual segmentation overestimates the lesion volume to ensure the entire lesion is identified [3] and the process is highly variable [4, 5]. A stable accurate segmentation is critical, as image features (such as texture and shape related features) are sensitive to small tumor boundary changes. Therefore, a highly automatic, accurate and reproducible lung tumor delineation algorithm would represent a significant advance.

Accurate extraction of soft tissue lesions from a given modality such as CT, PET or MRI is a topic of great interest for computer-aided diagnosis (CAD), computer-aided surgery, radiation treatment planning and medical research. However, segmentation of a lesion is typically a difficult task due to the large heterogeneity of cancer lesions (compared to normal tissues), noise that results from the image acquisition process and the characteristics of lesions often being very similar to those of the surrounding normal tissues. Traditional medical image segmentation techniques include intensity-based or morphological methods [6–9], yet these methods sometimes fail to provide accurate tumor segmentation.

A lung tumor analysis (LuTA) tool [10] within the Definiens Cognition Network Technology [11] was developed by Definiens AG [12] and Merck & Co., Inc. It is a prototype application that demonstrates the ability to automatically and semi-automatically identify and recognize organs and tumors in CT images. Its efficacy in automatic lung segmentation is described in [10] and we were able to obtain accurate lung segmentations from LUTA for all cases discussed here.

LuTA is designed to enable fast and easy annotation of lung tumors or other user-defined regions of interest. Flexible controls allow the annotation of structures of the user’s choice. Once a user has clicked on a region of interest, in a single two-dimensional (2D) slice, the application builds out the object three-dimensionally. The results [10] of the first prototype application built using the Definiens Cognition Network Technology for CT based scans provided a proof of concept enabling semi-automatic volumetric analysis of tumors. However, the current generation of the LuTA tool still has some drawbacks. First, although processing time is reduced compared to manual delineation, it still requires substantial operator input. For example, more than one user selected seed point may be required for tumor segmentation. Some tumor segmentations will not be completed in one step, thus additional operations are required (e.g. the radiologist must scroll over the CT slices and find out what part of the segmentation is missing). Additionally, lung tumor boundaries are often found to be incorrect during manual inspection, and thus require manual editing that takes additional time and creates additional sources of error. Finally, although segmentations can be performed in batch mode, which is appropriate for large studies, it is impractical if manual editing is required.

With the motivation of overcoming the above drawbacks of the “Click & grow” algorithm, we propose a new delineation algorithm based on using multiple seed points with region growing [13]. The new algorithm makes use of the original algorithm by using an original seed point to define an area, within which multiple seed points are automatically generated. An ensemble segmentation can be obtained from the multiple regions that were grown. Ensemble segmentation has played an important role in many medical image applications recently [14–17] and refers to a set of different input segmentations (multiple runs using the same segmentation technique with different initializations) that are combined in order to generate a consensus segmentation. In this paper, we demonstrate that such an approach reduces interobserver variability with significantly fewer operator interactions when compared to the original algorithm.

2. Related Work

More complex methods, such as energy minimization techniques, have been proposed and have been extensively applied in many studies within the last 10 years. Graph cut methods [18–24] and active contours (snake) [25–29] are two widely used methods that have been applied in many medical imaging applications. The graph cut method has been very popular in the area of image segmentation in recent years; it constructs an image-based graph and achieves a globally optimal solution of energy minimization functions. However, the biggest problem of the conventional graph cut algorithm is its computational cost, the running time and the memory consumption restricts its feasibility for many applications. In [24], a skeleton based-graph cut algorithm was introduced to more quickly classify volume data with high quality, extract important information about interesting structures, and decrease user interaction. A comparison with a skeleton based-graph cut algorithm is done in this paper. The active contours (snake) algorithm works similarly to a stretched elastic band being released. The initial points are defined around the object to be extracted. The points then move through an iterative process to a point with the lowest energy function value. The live wire or intelligent scissor [30–33] method is motivated by the general paradigm of the active contour algorithm, it changes the segmentation problem into an optimal graph search problem via local active contour analysis. Using dynamic programming the cost function is minimized. The live wire approach requires a user interactively define the boundary by moving a mouse along the region of interest, while the live wire process automatically computes a suggested boundary. Lu and Higgins [32, 33] recently proposed a single-section live wire based on a 2D section and a single-click live wire applied directly to 3D CT images to segment central-chest lymph nodes. The single-click live wire approach is similar to the single-section live wire, but is almost completely automatic. The single-click live wire idea is similar to the one we proposed here, only requiring one single seed in the desired location. Their method was applied to handle lymph node segmentation, which is quite different from lung tumor segmentation; it may be harder or easier than lung tumor segmentation depending on the lymph node location. Another extensively used method in recent years is the level set [34–39] algorithm. The level set method was first proposed by Osher and Sethian [39] in 1988 to track moving interfaces. The main idea behind the approach is to represent a contour as the zero level set of a higher dimensional function, called a level set function, and formulate the motion of the contour as the evolution of the level set function. A number of these approaches are finding commercial application. For example, the lesion sizing toolkit [38] is an open-source tool kit for CT lung lesions with integrated lesion sizing and level set algorithms. The CT lung lesion sizing tool is first used to detect four three-dimensional features corresponding to vasculature, the lung wall, lesion boundary edges and low density background lung parenchyma. Those features are the key to the segmentation process and they potentially prevent the segmentation from bleeding into non-lesion regions. Those features then were combined into a single feature using the feature aggregator method, the segmentation manager then applies a level-set region growing algorithm starting from a seed point and expanding until feature boundaries prohibit boundary advancement. This lesion sizing algorithm was added to ISP (Interactive Science Publishing) 2.3 [40], which is a 2D and 3D volume visualization application based on VolView [41]. ISP 2.3 is publicly available and was used in the current study to compare our results with the level set methods. In recently published works, statistical learning based approaches [42] [43] show us another way to handle segmentation problem. In Wu’s work [42], a system was created to mainly detect whether a lung nodule is attached to any of the major lung anatomies. The segmentation algorithm in their system played a very important role. It uses a conditional random field (CRF) model incorporating texture features, gray-level, shape, and edge cues to improve the segmentation of the nodule boundary. However, the purpose of their system at this segmentation stage is not to provide a perfect segmentation, but to apply a fast and robust method that can create a reasonable segmentation to serve as an input to a higher-level nodule connectivity classification system. Similarly, in Tao’s work [43], early detection of ground glass nodules (GGN) in lung CT images was presented, which is a multi-level statistical learning-based framework for automatic detection and segmentation of GGN. The system seems very promising. However, our work differs from above methods as we search for tumors.

In many medical image segmentation applications, one of the major drawbacks for most of the algorithms is that they require deliberate initialization, which becomes impractical when dealing with large numbers of images, this issue has been addressed in some of the existing works already. In Lu’s work [32, 33], a single click was required for live wire algorithm and Yan’s work [44] also shows it can find the minimal path for an object from a single starting point. One of our contributions in the paper is that the SCES requires only one initialization (one starting point).

3. Methods & Materials

3.1 LuTA Analysis Workflow

The overall goal for the LuTA implementation was to accurately, precisely and efficiently enable the analysis of lesions in the lung under the guidance of an operator. A standard analysis workflow was described in detail elsewhere [10]. The workflow is briefly described in the following:

A preprocessing step. This was designed to perform a segmentation of the lung as well as other off-line tasks, such as filtering, to improve the interactive performance of the analysis.

An optional step with semi-automated correction of the segmented lung. Since lesions are commonly found to be attached to the pleural surface, it was critical to enable efficient correction of the lung boundary in cases where the boundary between juxtapleural target lesions and the pleura had not been correctly determined during the automated preprocessing step.

A “Click&Grow” step with a user selected seed based segmentation of the lesions.

An optional manual refinement step of the semi-automated lesion segmentation to ensure medical expert agreement with any results that could influence patient management.

A reporting step generating volumes and statistics about other features, such as average density.

In this paper, we focused on how to accurately segment the lung lesion with minimum human interaction; the new algorithm is basically a substitution of steps 3 & 4 above and only requires one manual seed to be entered. The preprocessing and “Click&Grow” steps within the LuTA workflow are used by our new algorithm, as described below.

3.1.1 Preprocessing

The preprocessing step performs automated organ segmentation with the main goal of segmenting the aerated lung with correct identification of the pleural wall in order to facilitate the semi-automated segmentation of juxtapleural lesions. In Figure 1 , a CT image of a representative patient is shown segmented after the preprocessing step. The tumor is located in the right lung field.

Lung fields (left and right) were segmented after preprocessing.

3.1.2 Click & Grow

After the preprocessing step, the lung lesions must be located in one of the lung fields. In order to segment a target lesion the image analysts identified the lesion within the segmented lung and placed a seed point in its interior – typically at the perceived center of the lesion. Starting from the seed point, an initial seed object was automatically segmented using LuTa’s region growing based on similar intensities and proximity to areas with low intensity (“air”). This Definiens proprietary region growing process approximates the object’s surface tension T using an N 3 voxels sized kernel locally by calculating the ratio of the object volume ins >

From the grown region, which consists of voxels with similar density, the intensity weighted center of gravity (IWCOG) was calculated. To decrease inter and intra-reader variability, the seed point was shifted closer to the IWCOG. Additionally, an approximation of the lesion radius and volume, as well as histogram based lower and upper bounds for the intensity were extracted.

These parameters were used to define an octahedron-shaped candidate region within the lung. The new seed object was then grown into the candidate region with adaptive surface tension and intensity constraints.

The intensity constraints restrict the growth into candidate regions defined by: 1) a pre-computed intensity range of the Gaussian smoothed CT image, where the intensity range was estimated from the intensity statistics of the seed region and 2) a bound on the distance to the seed region which was calculated using a distance map. The distance map was calculated solely for the candidate region within the CNL local processing framework, and provides the minimal distance for each voxel to the seed region as an intensity value. Using the distance map ensures an approximate convexity of the seed object when growing into regions with similar intensities.

3.2 Single click ensemble segmentation

The original “Click&Grow” algorithm is very useful for delineating the tumor from the lung field in the LuTA application. If the growing process does not sufficiently capture the target lesion, the operator can place additional seed points within the lesion and repeat the growing process outlined above. Upon completion of the segmentation, the individual image objects are merged to form a single image object representing the segmented target lesion. The algorithm also provides the capability for a user to manually edit the segmentation. However, it still has the following drawbacks:

The segmentation is not consistent, different readers may generate different tumor regions.

It can require many human interactions (multiple clicks) to delineate the tumor in the case where the growing process did not sufficiently capture the target lesion.

The tumor boundary is often not satisfactory upon visual examination; sometimes it obviously includes many areas that do not belong to the target tumor.

To overcome the drawbacks of the original method, we propose a new algorithm: the single click ensemble segmentation (SCES) algorithm, which is an advanced version of the previous algorithm. The SCES makes use of the original algorithm by choosing different seed points automatically within a specified area of the lesion and performing region growing with each generated seed point. Thus, an ensemble segmentation is obtained from the multiple regions that were grown, and the final segmentation is based on a voting strategy. In order to better describe the algorithm, we provide first several definitions:

Tumor Core: the area most likely belonging to the tumor.

Manual Seed input: the first seed point provided by the user.

Start Seed point: the seed point randomly selected from the initial tumor region after shrinking.

Parent Seed point: algorithm selected seed point from a specific location of the Tumor Core.

Child Seed point: algorithm selected seed point from the outside of the existing tumor.

The detailed algorithm is described below and the detailed algorithm work flow is shown in Figure 2 :

My impression is that you only have BB and CC available right now. If that’s the case I’d just stick with the Cha-Ching until done. But really you don’t need either product, a standard bloom fertilizer like Tiger Bloom (2-8-4) should be sufficient when used with Big Bloom:
http://foxfarmfertilizer.com/index.php/item/tiger-bloom-liquid-plant-food.html

new grower Green o Magic and some clone unknow strain . Low experience.

smokerz

Auto Warrior

yesterday i use ff cha ching.

It early to use or not? ADVICE ME PLS.

And do u see problem on my plant.?

nomis

Auto Warrior

smokerz

Auto Warrior

i has flushed before give her cha ching 1/4 tsp for water 1.5litre.

last week i use bloom nutes it sell at my local start on 4week.

nomis

Auto Warrior

I’m still in the pseudo ‘new grower’ phase smokerz, with about 18 months of Fox Farm usage under my belt. So I’ll give you my opinion but maybe someone else more senior will stop in.

Your girl is quite small, and the main fan leaves are looking pale. I wonder if that is a result of transplanting or lack of light or other environmental issue?
I think you still need to run some veg nutes for another week or two at least, especially if you’re only at week 4 (I counted you up to week 6.5 so far though, did I make a mistake?). Do you have Grow Big/Big Bloom over there too? If not, what Vegging fertilizer do u have?

Cha-Ching is usually used the last 4 weeks (w8-12), and it’s also very powerful fertilizer supplement meant to be used in conjunction with standard bloom nutes like FF Tiger Bloom. I suppose if you have nothing else then it’s worth a try, but with your plant being so small I’m not sure you will see much of a benefit from using Cha-Ching. 1/4 tsp per 1.5 liter is too much (according to the instructions) and may cause a problem for you.

Overall the bud is looking decent and can see some trichs forming so think you’re in a good place still.
How are those clones doing?

smokerz

Auto Warrior

i think this fail begin at week 4 i add some bloom fertilizer for her(that time i not have any foxfarm fertilizer ). and i flush out before i use cha ching. yet my plant approx 6.5 week. and now i add little ff beastie (1/4tsp per 4 litre water) that way it better more than add cha-ching right? and when i start cha ching again.?

thank you so much for advice.
apologize for my poor English language.

nomis

Auto Warrior

Looking back over your grow I’m gonna guess that your girl got ‘stunted’ when you transplanted her. This is common for many auto flower plants, and many people recommend starting the seed in the final container and not bother with transplanting. Personally I germinate in a wet paper towel and then put the seed directly into my 5 gallon soil container.

You also appear to be using a perlite/vermiculite mix for soil, in the hempy grow style? But the last set of pictures looks like soil mix? I am intrigued as I’m looking to move into Hempy growing.
But if this is the case, then you most certainly will need to provide the base nutrients to the roots through the fertilizer in your water. It’s very much like hydro, so the water must be fresh and contain all the food your plant wants.

Here’s a link to the FF feed schedule, the supplements are listed at the bottom:
http://foxfarmfertilizer.comhttps://marijuanasaveslives.org/wp-content/uploads/pdf/SoilENG-Q2.pdf

Beastie Bloomz (0-50-30) contains no Nitrogen, whereas Cha Ching (9-50-10) has some N to keep things green at the end. Neither are meant to be used by themselves. They are meant to be used in conjunction with Grow Big/Big Bloom/Tiger Bloom or similar product line from another company. You may not be getting the ‘base nutrients’ you need for a healthy grow.

My impression is that you only have BB and CC available right now. If that’s the case I’d just stick with the Cha-Ching until done. But really you don’t need either product, a standard bloom fertilizer like Tiger Bloom (2-8-4) should be sufficient when used with Big Bloom:
http://foxfarmfertilizer.com/index.php/item/tiger-bloom-liquid-plant-food.html

Anyway dude your little girl looks ok, so for your first time I think you are doing pretty good there! If you can keep her healthy for a couple more weeks I think you’ll get a couple of decent buds off her.

:smoke: