Medical Imaging Segmentation

The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI) provides the largest set of computer tomography (CT) image references of lung nodules. The LIDC data set characterizes itself based on the following attributes.

• Calcification
• Internal structure
• Lobulation
• Malignancy
• Margin
• Sphericity
• Spiculation
• Subtlety
• Texture.

The annotated lesions in the LIDC IDRI are basically divided into three categories: “nodule> = 3 mm”, “nodule = 3 mm”. The Extensible Markup Language (XML) files accompanying the LIDC Digital Imaging and Communications in Medicine images contain the spatial locations of these three types of lesions

Lung cancer is most frequently diagnosed cancer. Computed tomography (CT) is an effective method and most common method for identifying lung cancer early. But, carefully examining each image from amongst a very large number of CT images, greatly inflates the burden of labor on radiologists. Also, radiologists tend to be subjective when using CT images for the diagnosis of lung disease, often leading to inconsistent results from the same radiologist at different times or from different radiologists examining the same CT image. To alleviate these diagnostic challenges, computer aided diagnosis systems, which use an automated image classification technique, can be used to help radiologists in terms of both their accuracy and speed.

The most commonly used measures for evaluating the usefulness of a new imaging modality are sensitivity (Se) and specificity (Sp). Sensitivity (True Positive Rate, TPR) and specificity (True Negative Rate) measure the ability of a test to correctly identify patient status as respectively diseased or non-diseased.

While segmentation is an important facet of medical imaging, it is fairly decisive to enhance the quality of the image. Median filter is used to remove any unwanted noise and frequency in an image.

Medical imaging indexing and retrieval is also important aspects with medical imaging. While many techniques (LBP and LTP) employ binary relations between the center pixel and its neighbors in 2Dlocal region of the image, some (DLTerQEP) employ the spatial relation between any pairs of neighbors along the given directions. DLTerQEP provide significant increase in discriminative power by allowing larger local pattern neighborhoods.

Medical Imaging have larger encoding length, meaning larger greyscale. When medical image is dealt in terms of superpixels - a group of pixels connected with similar greyscales, it would be forced to mark as a wrong label as it would contain larger number of pixels with close location which belong to both side of an edge. Algorithms are developed, which overcomes these issues by extending its neighbors to a larger area with more pixels.

Feature Extraction are typically on the outlines of diameter, volume, degree of roundness. Abundant techniques are available to extract features from a medical image. Once the features are extracted, the top most features which are optimal for classifying the lung nodules are listed and it will be classified based on those selected features.

One main reason for the difficulties in segmentation of the lung nodules is the attachment of other lung structures with the lung nodules. Automated corrected method can be applied to overcome this, which is done locally.

Many segmentation techniques are available, but can be mainly classified into Histogram based techniques, edge based techniques, region based techniques and Hybrid methods.

Segmentation of the lung nodules is a complex task. Sometimes a clearly visible lesion doesn’t have the information associated with it to declare it is a cancer tissue. Providing pixel wise probabilities will ignore all co-variances between pixels, making further analysis even tougher. Providing multiple hypothesis would benefit the pipeline of diagnoses treatments, which might lead to further diagnoses tests which resolves ambiguities. Most commonly an auto-encoder is used along with U-Net for segmentation of ambiguous medical images.

As lung segmentation is the preprocessing step before the lung detection, a region of interest (ROI) is generated for simplifying the segmentation process. Poor segmentation is often a performance drawback. Pulmonary nodules are generally classified into: isolated, juxtapleural, and juxtavascular. Isolated and juxtapleural are often found in the ROI and can easily be segmented. While juxtavascular can be missed. Semi-automated segmentation methods and bidirectional chain encoding methods are used to overcome missing of juxtavascular nodules. Whilst correcting the borders to avoid excluding nodules, over-segmentation needs to be minimized.

Convolutional neural networks can be used to learn the high-level representations from the training data. CNN along with the auto-encoders can be used to nodule classification. Thoracic CT produces a volume of slices that can be manipulated to demonstrate various volumetric representations of bodily structures in the lung. 3D CNN can is able to make full use of the 3D context information. 3D CNNs multi view strategy can achieve a lower error rate than the one-view-one-network strategy while using fewer parameters. Number of parameters, training time, and validation error rates needs to be considered when specifying which architecture is best suitable.

There exists an important class of images where even the full image context is not sufficient to resolve all ambiguities. Such ambiguities are common in medical imaging applications, e.g., in lung abnormalities segmentation from CT images. A lesion might be clearly visible, but the information about whether it is cancer tissue or not might not be available from this image alone.

In many case: Especially in medical applications where a subsequent diagnosis or a treatment depends on the segmentation map, an algorithm that only provides the most likely hypothesis might lead to misdiagnoses and sub-optimal treatment. Providing only pixel-wise probabilities ignores all co-variances between the pixels, which makes a subsequent analysis much more difficult if not impossible. If multiple consistent hypotheses are provided, these can be directly propagated into the next step in a diagnosis pipeline, they can be used to suggest further diagnostic tests to resolve the ambiguities, or an expert with access to additional information can select the appropriate one(s) for the subsequent steps.

While Computer aided detection (CADe) and Computer aided diagnoses (CADx) systems mostly operate independently, when integrated it into a single system best serves the identification and characterization of nodule. HOG-Histogram Oriented Gradient (description of features in an image), Water shed technique (Image Segmentation) when used together as a single system provides high accuracy and sensitivity.

There are many classification system for learning from and predicting the whole distribution of radiologists’ annotations. Such approach can be beneficial for CAD purposes for several reasons: learning from the distribution of annotations will help to avoid the loss of potentially important information when classification system has no knowledge on radiologists’ level of expertise.

Segmentation can be carried out on Fully Automated (FA) or Semi-automated (SA) or hybrid systems. While the FA systems require one or few control points, SA require more controls points which is a great deal of labor on user but the resulting system proves to be robust and can deal with challenging cases.

Luna16 helps to estimate the performance evaluation of different CAD systems, where each participant in this framework develops their algorithm and performs the tests on the same available dataset. The outcome of this challenge suggested that individual systems when combined yields better detection rate and sensitivity. The challenge consists of 2 separate tracks.

• complete nodule detection and
• False positive reduction.

While many technologies use pre-defined, hand-engineered feature models, a new technology where in custom radiomic sequencers is discovered that can generate radiomic sequences consisting of abstract imaging-based features tailored for characterizing lung tumor phenotype. Radiomics focuses on high-throughput extraction and analysis of a large number of imaging-based features for quantitative characterization and analysis of tumor. To implement concept of discovery radiomics for lung cancer detection, we introduce a deep convolutional radiomic sequencer that is discovered using a deep convolutional neural network. Since the custom radiomic sequencers are dependent on the data on which it is learned, the discovered radiomic sequencers produce highly tailor-made radiomic sequences for the tumor type, in this case lung lesions.