Calcification Characteristics and Clinical Visual Analysis System

The clinical visual analysis system that formed the basis for the design of our algorithm and guided our feature selection is described in detail elsewhere [2022]. It is based on several descriptors of the morphology and distribution of individual and clustered calcifications on mammograms. The number of calcifications in a cluster is not considered by itself a clear indicator of benign or malignant disease but when combined with other characteristics can increase or decrease suspiciousness [23-25]. The combination of all these properties by the human observer was shown to yield a sensitivity of 97.6% (correct identification of cancers associated with calcifications) and a specificity of 73.3% (correct identification of benign cases associated with calcifications) [20, 22].

The Breast Imaging Reporting and Data System (BIRADS) Lexicon of the American College of Radiology (ACR) was established in 1993 in an effort to standardize and improve mammographic interpretation. BIRADS was based on the clinical visual system of analysis. The recommended BIRADS descriptors for calcifications and calcification clusters are summarized in Table 13.1 [26]. Overall, there is strong evidence that morphology and distribution are two of the most important clinical aids in making the diagnosis of mammographic calcifications.

In clinical practice, a radiologist makes the final diagnosis of the detected calcifications based on the BIRADS characteristics, demographic information, and associated mammographic findings. However, inter- and intraobserver variability in the assignment of morphological features to the identified calcifications and ambiguity in the interpretation significantly degrades diagnostic performance. Hence, successful differentiation is limited among radiologists and can be as low as 20% leading to numerous unnecessary biopsies of cases with calcifications clusters [27].

Computer algorithms could translate and automate the clinical experience and thus assist the radiologist in this diagnostic task. An algorithm that provides information on the morphology, e.g., segments calcifications while preserving size and shape, and gives a likelihood of malignancy for a detected calcification cluster could be extremely valuable in mammogram interpretation and patient

Table 13.1: BIRADS descriptors for calcifications with associated genesis type [26]

Morphology or Skin (lucent centered) B

character Vascular (linear tubular with parallel tracks) B

Coarse or popcorn like B

Large rod-like B

Round (larger than 0.5 mm) B

Eggshell or rim (thin walled lucent centered, cystic) B

Milk of calcium (varying appearance in projections) B

Dystrophic (irregular in shape, over 0.5 mm, lucent centered) B

Punctate (round smaller than 0.5 mm) B

Suture (linear or tubular, with knots) B

Spherical or lucent center (smooth and round or oval) B

Amorphous or indistinct U

Pleomorphic or heterogeneous granular M

Fine linear M

Fine linear branching M

Distribution Clustered U

Segmental U/M

Regional U

Diffuse/Scattered B

Linear M

Number 1-5 U

B = probably benign; M = suggestive of malignancy; U = uncertain.

management. To be clinically useful, the algorithm should perform real time, be robust, and have consistent performance at least comparable to the clinical visual analysis system [20, 22]. The algorithm described here was designed to meet the above requirements and two additional conditions: (a) The desired classification performance had to be achieved with the smallest possible set of features. (b) Feature selection would be initially limited to the morphological, distributional, and demographic domains; expansion to other domains would be considered only if performance did not reach desirable levels. The specific components of this scheme are shown in Fig. 13.3. The algorithm was implemented and tested on simulated calcification clusters, large sets of mammographic calcifications, and datasets of various image resolutions [20, 28, 29]. All studies demonstrated that the development of a classifier on morphological characteristics alone is a viable and reliable approach. They also supported our hypothesis

Digitized Image

Classification ANN

Leave-one-out resampling

ROIs 512x512

Feature Definition

Thresholding

Detection/ Segmentation iz

Shape Analysis

Likelihood of malignancy

Figure 13.3: Flowchart of the CADiagnosis algorithm developed for the differentiation of benign from malignant microcalcification clusters in digitized, screen/film mammography [20].

that a classifier could be used as an indirect measure of segmentation performance [29]. The segmentation of the individual calcifications and the clusters with shape and distribution preservation was a critical stage in our methodology. Hence, in the following section, we will discuss the detection/segmentation stage in more detail with particular emphasis on the role of wavelets in this process.

0 0

Post a comment