Development of computer-aided detection methods for CMBs is challenging because of their small size, wide distribution, and the presence of other structures that mimic their appearance. The need to identify radiation-induced CMBs with low contrast is a further factor that complicates their detection. Despite several previous attempts to address this problem, it remains a topic of active research. In this study, we have presented a semi-automated CMB detection algorithm that has achieved superior performance with a sensitivity of 86.5% and computation time of 1 min.
Achieving high sensitivity was the top priority in designing our algorithm because it has more impact than specificity on the performance that can be achieved. From a practical standpoint, visual inspection to remove false positive lesions is far easier than identifying CMBs missed by the detection algorithm, as the eye can more rapidly dismiss lesions falsely labeled as CMBs than it can search the entire imaging volume for true CMBs. The overall performance of our algorithm depended on both the size and contrast of CMBs, which is supported by the heightened sensitivity that was observed for definite CMBs compared to possible ones (95.4% vs. 77.5%). The newly formed radiation-induced CMBs that were evaluated in this study are typically harder to detect than CMBs caused by trauma or other mechanisms of disease, as they are often smaller in size and have varied contrast. Thus, we anticipate that our algorithm would be even more robust when applied to other patient populations.
The high sensitivity of our algorithm may be explained to the strength of our false positive reduction strategy. Screening of the FRST map by applying a vessel mask allowed the use of a much lower threshold for the FRST map and maximized sensitivity to smaller and low contrast CMBs during the initial detection step. This method facilitated the detection of 96.1% of all true CMBs, which included 100% of definite CMBs and missed a relatively small number of possible low-contrast CMBs. If this vessel mask had not been applied, it would have been necessary to use a higher threshold, which would have sacrificed the detection sensitivity. Although true CMBs can be masked if they are adjacent to the structures that are included, the number of these CMBs was small (2.0%; 1 definite and 5 possible) for our experiments.
Another significant contributor to our high sensitivity in initial detection was the 3D region growing process. Although a potential pitfall that is inherent for region growing is the occasional inclusion of neighboring pixels containing nearby vessels or susceptibility artifacts (c), we only applied region growing to a subset of CMBs with lower contrast and smaller size. Bypassing region growing for large CMB candidates afforded reduced thresholds for geometric features, which lead to more efficient false positive reduction. This separate processing pipeline for large and small CMBs resulted in the removal of only 7.6% (6 definite and 17 possible) of true CMBs during subsequent region growing.
Our method achieved a higher sensitivity, faster computation speed, and a reduced number of false positives compared with other related approaches. Also the performance of our algorithm was validated on a larger sample size in terms of total number of CMBs and incidence per patient (), demonstrating the robustness of our method. In particular there was heightened sensitivity (86.5% vs. 71.2%) compared to the other method based on FRST (Kuijf et al., 2012
). Moreover, it should be noted that many of our CMBs spanned a smaller number of pixels than Kuijf et al. (2012)
(average diameter: 2.08 vs. 2.28 pixels) and hence posed a greater challenge. This improved sensitivity applied even to the possible
CMBs that had an average diameter of 1.84 pixels and after normalization had an average minimum intensity up to 26.8% of the maximum.
Unlike previously published studies, our elevated sensitivity is achieved without compromising computation speed. The fast computation of our algorithm originates from its simple design. Skull extraction and intensity normalization are the only preprocessing steps required, whereas other methods utilize image registration (Kuijf et al., 2012; Seghier et al., 2011
) and/or segmentation (Kuijf et al., 2012
) routines prior to applying their detection algorithms. In the initial detection step, gradient-based FRST can quickly locate local hypointense regions as potential CMBs. The computation time for FRST is further reduced in our algorithm by eliminating the convolution step originally proposed by Loy and Zelinsky (2003)
. In addition, the transform is repeatedly computed only at 3 radii in our experiments, while it was compute by Kuijf et al. at 18 radii. Finally, 3D region growing is performed within only a small local region for each candidate CMB, and only 3 geometrical features (area, circularity and centroid) are quantified to sequentially remove falsely identified CMBs, whereas the method of Barnes et al. (2011)
utilizes 14 features associated with the shape, intensity, and size of CMBs in a support vector machine to perform the classification, which takes up to multiple days for parameter training.
3D region growing was implemented in our algorithm in order to reduce the number of false positives that were present after the FRST step. This allowed geometric features of a potential CMB to be extracted and was used in our study to eliminate 86.1% of false positives. The remaining false positives (1.5/CMB and 44.9/patient) originated mostly from vessels, susceptibility artifacts, or the surgical cavity. Overall, our algorithm produced a smaller percentage of false positives per true CMB than previous ones. Since radiation-induced CMBs are often smaller and of lower contrast, there is a trade-off between producing a small number of false positives and maintaining the high detection sensitivity achieved in our study. Kuijf et al. (2012)
used gray/white matter masks to exclude brain structures such as ventricles and sulci, where false positives are often observed. The disadvantage of this approach is that it requires the acquisition of a T1-weighted image as well as the application of registration and segmentation algorithms that prolong the total processing time. In addition, all of our patients received intracranial tumor resection, which produces structures that mimic CMB (see b). These structures can be removed quickly during the final visual inspection because of their obvious anatomical location. A strict performance comparison of automated detection algorithms is difficult at this point, as the size and contrast of CMBs may vary with other experimental factors such as field strength and resolution (Nandigam et al., 2009
). The manual review process and patient inclusion criteria are also all different among these studies. It is desirable in the future to construct a standard CMB database, in which CMBs are categorized by their disease type, MR imaging field strength, distribution or other related factors. This will not only help objectively evaluate automated CMB detection algorithms but also facilitate the training process for these algorithms.
The success of our algorithm demonstrates the advantage of a using a vessel mask to remove false positives in achieving a high sensitivity while maintaining a reasonable specificity in CMB detection. While methods that are able to create a high quality vessel mask in 3D using SWI have been proposed in literature (Koopmans et al., 2008
), we used the vessel mask from the FRST transform in our approach because of its robustness and simplicity in integrating with our algorithm's pipeline. The better delineation and continuity of veins on mIP SWI compared to non-projected images facilitated both the generation of a more reliable vessel mask by FRST and a greater extent of region growing on these structures, both of which aided in reducing false positives. The improved false positive reduction can in turn be used to enhance detection sensitivity, e.g., set up a low threshold to screen the FRST map and make the initial detection highly sensitive. Limitations associated with the usage of mIP SWI images include susceptibility artifacts at air–tissue interfaces and magnified background noise that is introduced during reconstruction of SWI images that can be a potential source of false positives (c). Also the original location of CMBs cannot be determined on these images because of the projection processing, but it can easily be recovered on the original non-projected images. The projection processing may also accidentally project some CMBs in or close to certain dark structures that do not surround them in actual anatomy, leading to decreased FRST response or leakage in region growing, which both increase the number of false negatives. Despite these limitations, using FRST on mIP SWI images is especially advantageous for the detection of radiation-induced CMBs, whose essentially low contrast on conventional T2*-weighted magnitude images as the result of its small size is greatly improved on mIP SWI (Lupo et al., 2012
). Furthermore, processing MR image using mIP has been suggested as a required step for visual CMB detection, especially when high resolution images acquired at high field strengths (e.g., 3 T or 7 T) are used (de Bresser et al., in press
). As visual inspection (e.g., further false positive removal) of the output from automated CMB detection is still required, it is desirable to design an algorithm that operates on the same images that are used for subjective interpretation.
Finally, there are several limitations associated with our algorithm itself. First, while 3D region growing is a simple and fast method to segment putative CMBs, its capability to discriminate desired objects from close but dissimilar background is limited, making the segmentation leakage being more likely to happen for CMBs with low image contrast. To improve the accuracy of the segmentation, more advanced and finer techniques such as active contours and let sets may be used (Osher and Sethian, 1988
), but at the expense of computation time. Second, at the very first and last image slices, the efficiency of geometric examination becomes lower due to limited space for region growing along slice selection direction. This may result in there being a higher number of false positives on these slices than on inner slices. Like previous published methods, training is an inevitable step that is required to apply our algorithm to detect CMBs from different types of diseases, different MR field strengths, and even different scan parameters such as TE, spatial resolution, and slice thickness. Size and contrast variation are the principal reasons that training is necessary, as most of parameters selected in our algorithm require that the range of CMB size and intensity be considered. However, the time used for training our algorithm is small as most of the parameters used in the algorithm can be empirically determined by studying a few representative CMBs as long as prior knowledge about the size of CMBs is available. Optimal values for other parameters such as MID and C can be determined empirically or formally by constructing ROC curves on a small training dataset such as we did in this study.