I-Jen Chiang

Automatic recognition of midline shift on brain CT images

Midline shift is one of the most important quantitative features clinicians use to evaluate the severity of brain compression by various pathologies. It can be recognized by modeling brain deformation according to the estimated biomechanical properties of the brain and the cerebrospinal fluid spaces. This paper proposes a novel method to identify the deformed midline according to the above hypothesis. In this model, the deformed midline is decomposed into three segments: the upper and the lower straight segments representing parts of the tough dura mater separating two brain hemispheres, and the central curved segment formed by a quadratic Bezier curve, representing the intervening soft brain tissue. The deformed midline is obtained by minimizing the summed square of the differences across all midline pixels, to simulate maximal bilateral symmetry. A genetic algorithm is applied to derive the optimal values of the control points of the Bezier curve. Our algorithm was evaluated on pathological images from 81 consecutive patients treated in a single institute over a period of one year. Our algorithm is able to recognize the deformed midlines in 65 (80%) of the patients with an accuracy of 95%, making it a useful tool for clinical decision-making.

A simple genetic algorithm for tracing the deformed midline on a single slice of brain CT using quadratic Bezier curves

Midline shift (MLS) is one of the most important quantitative features clinicians use to evaluate the severity of brain compression. It can be recognized by modeling brain deformation according to the estimated biomechanical properties of the brain structures. This paper proposes a novel method to identify the deformed midline by decomposing it into three segments: the upper and the lower straight segments representing parts of the tough meninges separating two brain hemispheres, and the central curved segment formed by a quadratic Bezier curve, representing the intervening soft brain tissue. The deformed midline is obtained by minimizing the summed square of the differences across all midline points, applying a genetic algorithm. Our algorithm was evaluated on images containing various pathologies from 81 consecutive patients treated in a single institute over one-year period. The deformed midlines were evaluated by human experts, and the values of midline shift were accurate in 95%

TRACING THE DEFORMED MIDLINE ON BRAIN CT

Midline shift (MLS) is the most important quantitative feature clinicians use to evaluate the severity of brain compression by various pathologies. We proposed a model of the deformed midline according to the biomechanical properties of different types of intracranial tissue. The model comprised three segments. The upper and lower straight segments represented parts of the tough meninges separating two hemispheres, and the central curved segment, formed by a quadratic Bezier curve, represented the intervening soft brain tissue. For each point of the model, the intensity difference was calculated over 48 adjacent point pairs at each side. The deformed midline was considered ideal as summed square of the difference across all midline points approaches global minimum, simulating maximal bilateral symmetry. Genetic algorithm was applied to optimize the values of the three control points of the Bezier curve. Our system was tested on images containing various pathologies from 81 consecutive patients treated in a single institute over one-year period. The deformed midlines itself as well as the amount of midline shift were evaluated by human experts, with satisfactory results.

Automatic MRI Meningioma Segmentation Using Estimation Maximization

With the advancement of the imaging facility and image processing technique, computer assisted surgical planning and image guided technology have become increasingly used in neurosurgery. For MRI has the characteristic of multi-spectral image data, so knowledge-base techniques is widely used in brain MRI segmentation. Here we recognize the location of the tumor automatically and provide an accurate result by estimation maximization method. Simultaneously, promote the efficiency of reading image as well

A knowledge discovery approach to diagnosing intracranial hematomas on brain CT: recognition, measurement and classification

Computed tomography (CT) of the brain is preferred study on neurological emergencies. Physicians use CT to diagnose various types of intracranial hematomas, including epidural, subdural and intracerebral hematomas according to their locations and shapes. We propose a novel method that can automatically diagnose intracranial hematomas by combining machine vision and knowledge discovery techniques. The skull on the CT slice is located and the depth of each intracranial pixel is labeled. After normalization of the pixel intensities by their depth, the hyperdense area of intracranial hematoma is segmented with multi-resolution thresholding and region-growing. We then apply C4.5 algorithm to construct a decision tree using the features of the segmented hematoma and the diagnoses made by physicians. The algorithm was evaluated on 48 pathological images treated in a single institute. The two discovered rules closely resemble those used by human experts, and are able to make correct diagnoses in all cases.

Discovering Latent Semantics in Web Documents Using Fuzzy Clustering

Web documents are heterogeneous and complex. There exists complicated associations within one web document and linking to the others. The high interactions between terms in documents demonstrate vague and ambiguous meanings. Efficient and effective clustering methods to discover latent and coherent meanings in context are necessary. This paper presents a fuzzy linguistic topological space along with a fuzzy clustering algorithm to discover the contextual meaning in the web documents. The proposed algorithm extracts features from the web documents using conditional random field methods and builds a fuzzy linguistic topological space based on the associations of features. The associations of cooccurring features organize a hierarchy of connected semantic complexes called “CONCEPTS,” wherein a fuzzy linguistic measure is applied on each complex to evaluate 1) the relevance of a document belonging to a topic, and 2) the difference between the other topics. Web contents are able to be clustered into topics in the hierarchy depending on their fuzzy linguistic measures; web users can further explore the CONCEPTS of web contents accordingly. Besides the algorithm applicability in web text domains, it can be extended to other applications, such as data mining, bioinformatics, content-based, or collaborative information filtering, etc.

A multiresolution binary level set method and its application to intracranial hematoma segmentation

We propose a multiresolution binary level set method for image segmentation. The binary level set formulation is based on the Song-Chan algorithm, which cannot compute the edge length when the margin of the image is irregular. We modify the edge length approximation so that it can work everywhere in a single-connected image, make it suitable to segment objects at any position, especially near the margin of the image. For multiresolution processing, we use image pyramids. The binary level set method works on images with reduced resolution and size. A point at the image with lower resolution is processed instead of a block or a strip at the original resolution, therefore improving the efficiency. Our multiresolution binary level set method is applied to segmentation of intracranial hematomas on brain CT slices. Segmentation of epidural and subdural hematomas, which have been not done previously, is performed successfully in seconds with results comparable to human experts.

PICO element detection in medical text without metadata

Efficient identification of patient, intervention, comparison, and outcome (PICO) components in medical articles is helpful in evidence-based medicine. The purpose of this study is to clarify whether first sentences of these components are good enough to train naive Bayes classifiers for sentence-level PICO element detection. We extracted 19,854 structured abstracts of randomized controlled trials with any P/I/O label from PubMed for naive Bayes classifiers training. Performances of classifiers trained by first sentences of each section (CF) and those trained by all sentences (CA) were compared using all sentences by ten-fold cross-validation. The results measured by recall, precision, and F-measures show that there are no significant differences in performance between CF and CA for detection of O-element (F-measure=0.731±0.009 vs. 0.738±0.010, p=0.123). However, CA perform better for I-elements, in terms of recall (0.752±0.012 vs. 0.620±0.007, p<0.001) and F-measures (0.728±0.006 vs. 0.662±0.007, p<0.001). For P-elements, CF have higher precision (0.714±0.009 vs. 0.665±0.010, p<0.001), but lower recall (0.766±0.013 vs. 0.811±0.012, p<0.001). CF are not always better than CA in sentence-level PICO element detection. Their performance varies in detecting different elements.

Estimating postoperative skull defect volume from CT images using the ABC method.

OBJECTIVES Surgeons often perform decompressive craniectomy to alleviate a medically-refractory increase of intracranial pressure. The frequency of this type of surgery is on the rise. The goal of this study is to develop a simple formula for clinicians to estimate the volume of the skull defect, based on postoperative computed tomography (CT) studies. METHODS We collected thirty sets of postoperative CT images from patients undergoing craniectomy. We measured the skull defect volume by computer-assisted volumetric analysis (V(m)) and our own ABC technique (V(abc)). We then compared the volumes measured by these two methods. RESULTS The V(m) ranged from 3.2 to 76.4 mL, with a mean of 38.9 mL. The V(abc) ranged from 3.8 to 71.5 mL, with a mean of 38.5 mL. The absolute differences between V(abc) and V(m) ranged from 0.05 to 17.5 mL (mean: 3.8±4.2). There was no statistically significant difference between V(abc) and V(m) (p=0.961). The correlation coefficient between V(abc) and V(m) was 0.969. In linear regression analysis, the slope was 1.00086 and the intercept was -0.0035 mL (r(2)=0.939). The residual was 5.7 mL. CONCLUSION We confirmed that the ABC technique is a simple and accurate method for estimating skull defect volume, and we recommend routine application of this formula for all decompressive craniectomies.

Automatic measurement of midline shift on deformed brains using multiresolution binary level set method and Hough transform

Midline shift (MLS) is an important quantitative feature clinicians use to evaluate the severity of brain compression by various pathologies. The midline consists of many anatomical structures including the septum pellucidum (SP), a thin membrane between the frontal horns (FH) of the lateral ventricles. We proposed a procedure that can measure MLS by recognizing the SP within the given CT study. The FH region is selected from all ventricular regions by expert rules and the multiresolution binary level set method. The SP is recognized using Hough transform, weighted by repeated morphological erosion. Our system is tested on images from 80 patients admitted to the neurosurgical intensive care unit. The results are evaluated by human experts. The mean difference between automatic and manual MLS measurements is 0.23 ± 0.52 mm. Our method is robust and can be applied in emergency and routine settings.