Abstract
In this work, we describe a novel symbolic representation of shapes for quantifying skull abnormalities in children with craniosynostosis. We show the efficacy of our work by demonstrating an application of this representation in shape-based retrieval of skull morphologies. This tool will enable correlation with potential pathogenesis and prognosis in order to enhance medical care.
Craniosynostosis, the premature fusion of the fibrous skull joints, or sutures, is a serious and common pediatric disease. Depending on the location of the suture fusion, various abnormal head shapes may result. These shapes may be correlated with predisposing or causative environmental and genetic factors and resultant deficits in neurocognition. Currently, interpretation of these images remains largely confined to subjective description with high interobserver variability and poor reproducibility. To advance the study of craniosynostosis, it is critical to develop quantitative shape descriptors that will enable identification of correlations with genotype and neurocognition. In this poster, we describe a symbolic representation of shapes that allows us to quantify skull shape for applications in shape-based retrieval of skull-morphologies.
In general document retrieval, key words are used to search for similar documents because different subjects often possess unique terms that are different from others. In our work, the words translate into features, and the documents into different skull shapes. Each skull shape is considered as a subject and may possess a unique pattern of features.
Representative 2-D sections are selected from 3-D volumetric head CT scans of 60 sagittal and 13 metopic synostosis patients and 40 controls. Starting with a set of equally-spaced vertices around the contour of each image, a vector of distances from each vertex to all the others is computed. These vectors are clustered by the K-means algorithm and each cluster assigned a letter as its label. Three letters from each sequence of three adjacent vertices form words, and the skull is represented by a set of these words. Figure 2 shows an output of word frequencies used in all patients in the study. As shown in the figure, each type of head shape is described by words and term-frequencies that are different from other shapes.
Figure 2.
Frequency of different words used in each document. X-axis represents words used in all documents.
To show the efficacy of our work, a test set of 40 sagittal and metopic images was used to find the best matching shapes in our database. These queries were done using the words in each image. Of the 40 queries made, only 1 had a false match. Figure 3 shows an example of a metopic query and 3 of the top matches from the symbolic shape-based retrieval.
Figure 3.
Three images on the right are the best matching shapes from the metopic query (leftmost image) in our database.
With the high accuracy shape stratification, the proposed novel shape descriptor may be an important tool for advancing our understanding of correlations between shapes and pathogenesis and prognosis of developmental diseases. In addition, techniques and tools that are developed in general document retrieval may now be used on skull images.
Figure 1.
Left: Normal skull with various sutures. Middle: Elongated skull shape representative of sagittal synostosis. Right: Triangular skull shape representative of metopic synotosis.