• No results found

Knowledge-based search methods

5.2 Existing preprocessing and region of interest extraction methods

5.2.3 Contrast enhancement

5.2.4.2 Knowledge-based search methods

The majority of existing methods for ROI extraction are based on knowledge-based search methods. These are heuristic search methods that use prior knowledge of the soft tissue or bone anatomy of the hand and wrist. They are usually implemented as a series of succes- sive refinement operations, starting with identification of the fingers or radius and ulna, moving on to bone identification, then finishing with delineation of individual regions of interest. Sometimes the individual regions require further refinement like correction for radiograph misalignment so they are in a standard orientation ready for further analy- sis [Piet04].

A common approach to initial identification of regions is to threshold the image using ei- ther a single threshold value [Mich89] [Piet91] [Hsie07b], or adaptive thresholding method [Luis03] [Zhan06] [Zhan07]. This produces a silhouette of the hand soft tissue compo- nent, but it may include some bone outline components, especially for the distal pha- langes [Piet97]. The silhouette can be cleaned-up using the morphological operations of dilation and erosion to remove unwanted regions (for example, [Effo94]).

From the silhouette image, a number of techniques have been used to locate and label the major anatomical regions of the hand. For the most part, these techniques can be grouped into line profile search methods, and methods for identifying angle changes of the traced hand contour. The profile search methods scan across the hand and run-length encode step changes between black and white values in the binary image, or some other measure of change. These step changes are related to key anatomical landmarks like the edges of the fingers. By aligning the centre of the steps, it is possible to identify, for example, the phalangeal axes [Sun94] [Piet01b] [Hill94a] [Effo94]. After smoothing the image profiles within the steps, derivative operations can be used to find bone edges for subsequent pro- cessing [Mich89].

The angle-change method works by tracing the outline of the silhouette and looking for changes in contour angle with distance along the contour. Contour angle changes, like those at the tips of the phalanges, can be used to identify anatomical landmarks [Effo94] [Mahm00] [Kwab85] [Lee08] [Yoo07]. Another use of tracing the contour is to sweep across an arc, analysing the distance from a fixed point in the wrist to the silhouette outline, and looking for distance versus angle patterns that correspond to anatomical landmarks, like valleys corresponding to the short distance to the base of the phalanges [Mahm98].

The silhouette can also be skeletonised using a thinning algorithm or medial axis trans- form. This process results in axes that are first approximations to the centre lines for the phalanges [Gabo97] [Luis03]. A collection of search paths are created by linking these

5.2 Existing preprocessing and region of interest extraction methods 103

axes and extending them to the image boundary or soft tissue-background margin. By analysing projections perpendicular to these search paths, it is possible to identify joints and extract phalangeal EMROIs [Gabo97]. However, it is not possible to find the metacarpals using this method. They have been found instead by using a circumferential search across the hand at a fixed distance along the extended axes [Luis03].

One problem with the silhouette method is that it is reliant upon the initial thresholding or segmentation of the soft tissue and bone areas. With variable image background and soft tissue thickness variations and overlap (Figure 5.4(a)), standard segmentation methods begin to fail. An alternative is to either find ridges in the image corresponding to the long axes of the fingers [Marq01], or to find bone edges using edge detectors like the Sobel filter [Piet91] [Morr94] and the Canny edge detector [Mano94] [Hsie07b].

The carpal region of interest (CROI) can be found by scanning vertical and horizontal lines of the silhouette and looking for edges and widths that relate to anatomical features like the minimum width of the proximal end of the wrist, and the soft tissue junction between the thumb and second digit [Piet93] [Zhan07]. Because fixed horizontal and vertical scan lines are used in this method, the CROI results will depend on the performance of the orientation correction method. To some degree, this orientation limitation can be overcome by using a method that starts by finding a third metacarpal reference axis, then uses empirical CROI ratios and skin boundary regression lines to extract the CROI [Voge00].

Instead of using the silhouette to find soft tissue landmarks, it is possible to delineate a CROI by finding bone landmarks. This will only be reliable if the bones are present across all stages of maturity. One set of landmarks that has been used is the head of the ulna and the base of the thumb [Hsie07b]. These landmarks can be found by scanning the image to find the head of the ulna bone and projecting across the bone to define the base of the CROI, then scanning to find the base of thumb and projecting across the first metacarpal to define the top of the CROI [Hsie07b].

Finally, in a method referred to as the ‘sticking’ algorithm, a simulated needle is stuck side- ways into the hand image and a histogram of pixel intensities tracked until local thresholds have been found corresponding to the bone edges [Fan01]. This method can be used to find upper and lower bounds for the CROI. Using these bounds, the rest of the carpus can be divided into bone search regions using projective pixel-intensity integration operations. This method requires good contrast between bone and soft tissue, and only works in chil- dren up to 7 years old because after this age the carpal bone separation is too small and the extraction error increases [Fan01].