Image pre-processing is essential; it normalizes data for a particular method. One worth mentioning fact is that digital mammograms (as those used in our project) greatly abandon many preprocessing steps. It is evident that those unpleasing artifacts, marks, and image noise in the analogue image have no existence in the digital mammogram; this resulted in a much cleaner image.
After noise removal, breast region is extracted from the image; this step is called segmentation. Then, the image is normalized to make sure that images from different mammography machines lay on the same scale;this casts LIBCAD as a mammography machine independent
The final steps are detecting the microcalcification foci, and then grouping those foci into different clusters. The radiologist can opt to view only dense clusters by manipulating a threshold level. CAD detects and marks microcalcification foci, even if those foci that are not clustered. To measure the accuracy of the algorithm, we define a cluster as follows. It is a set of detected foci, where any two of them are at most 3 mm apart. If the number of foci per cluster is lower than the selected threshold level, the cluster is considered undetected although those foci will be marked to the radiologist. A cluster is considered a Positive cluster, and hence is counted, if the number of its detected foci is larger than the selected threshold level and its centre is located within the true marking (ground truth) of the radiologist proven by a biopsy.
Clinical Trial is the most important measuring index. It is used to compare the performance of radiologists with and without the use of the CAD. This is because CADs are not intended for fully automatic usage; rather, they are “aiding” radiologists. The final decision is the responsibility of the radiologist.Therefore, the right performance index is to measure the relative performance of the radiologist with and without using the CAD.
This involves designing the trial, interacting with radiologists to train them how to use the CAD, and doing statistical analysis for the results. It is worth mentioning that for any CAD to be approved by the FDA a clinical trial must be conducted. MESC Labs started the process of clinical trial design. The figure, below, is a snapshot from our software designed for this clinical trial
Snapshot of our software designed to be used in clinical trials
The working team comprises a multidisciplinary group of several backgrounds including statistics, computer science, and engineering, along with a trained, experienced and professional radiologist (10 years’ experience, 6000 mammogram/year). This is in addition to collaboration with partners from industry and marketing.
Mammograms are collected from two different institutions. All images are acquired from digital mammography. The radiologist reads the digital mammograms and then marks the lesions in the images. The marked lesions are also tagged according to the different radiological lexicons and then categorized by the radiologist according to the ‘‘Breast Imaging Reporting and Data System’’ (BIRADS) scoring system.
We used 100 normal cases (437 images) to calculate the False Positive (FP) results and 488 cases (1952 images) with abnormalities. All the images are digital mammography. Out of these 488 cases, only 38 cases (67 images) have malignant microcalcifications. Those 38 cases are used to calculate the True Positive findings (sensitivity).
Malignant microcalcifications were detected by the radiologist in 100% (38/38) of cases: 86.8% (33/38) microcalcifications alone and 13.2% (5/38) microcalcifications with masses. The performance was tested at two threshold levels. At a threshold of 4 foci per cluster (an aggressive threshold) malignant microcalcifications were detected in 97.4% (37/38) of cases: 86.8% (33/38) microcalcifications alone and 10.5% (4/38) microcalcifications with masses. At a threshold of 8 foci per cluster (a less aggressive threshold) the detection rate was 92.1% (35/38) of cases: 84.2% (32/38) microcalcifications alone, and 7.9% (3/38) microcalcifications with masses.
We have implemented our protocol for reading and marking by designing software that facilitates the radiologist with labelling, marking, and attributing tools. All lesions were classified according to the BIRADS system, then a BIRADS score was assigned for each image (10). All suspicious lesions classified as BIRADS 3, 4 and 5 are pathologically proven after core and vacuum needle biopsy.
1- Abdel Razek, N. M., Yousef W. A., and W. A. Mustafa (2013). "Microcalcification detection with and without CAD system (LIBCAD): A comparative study." The Egyptian Journal of Radiology and Nuclear Medicine 44(2): 397-404.
2- Abdelrazek, N.; Yousef, W.; Mustafa, W. (2012), "Microcalcification detection with and without prototype CAD system (LIBCAD): a comparative study", European Society of Radiology (ECR 2012 / C-1063).
3- Yousef, W. A. et al. (2010), ``On Detecting Abnormalities in Digital Mammography''. in Applied Imagery Pattern Recognition Workshop 2010. Proceedings. 39th; IEEE Computer Society.