Automated Detection of Follicular Centers for Follicular Lymphoma Grading
Jun Kong ; Department of Biomedical Informatics, The Ohio State University; Kim Boyer ; Department of Electrical and Computer Engineering, The Ohio State University; Gerard Lozanski ; Department of Pathology, The Ohio State University; Olcay Sertel ; Department of Biomedical Informatics, The Ohio State University;
Content:
Follicular Lymphoma (FL) originates from follicle center B cells and is comprised of centrocytes and centroblasts. Diagnosis of FL is based on characteristic morphologic, immunophenotypic and cytogenetic abnormalities. Histological grading of FL is based on the number of large malignant centroblasts. Centroblast count is performed in ten different and representative neoplastic follicles using an ocular with an 18mm field of view at 1x magnification and a 40x dry objective resulting in a high power field of 0.159 mm2. Based on this method FL is stratified into three histological grades. Unfortunately, the manual method of grading follicular lymphomas is difficult, time consuming and prone to poor reproducibility. An attractive solution to this problem is to develop a computerized system for histologic grading of FL.
Technology:
The inputs to our system are images of tumor slides stained with haematoxylin and eosin, and digitized with the ScanScope T2 digitizer (Aperio, San Diego, CA). The developed method makes full use of the discriminative information derived from the cell spatial density to identify the boundaries between follicular centers and their surrounding mantle zones. A sequence of computer vision and image analysis steps, including color feature clustering, feature measures extracted from cell spatial density and morphological operations, are followed in the proposed method. As the output, the boundary of the resulting detected follicle center is represented and smoothed with the Fourier descriptor in the post-processing step.
Design:
Since the follicular patterns are very sensitive to the magnification scale with which they are presented, we chose to work on images at 2x magnification where the follicular pattern is most distinctive. The accuracy was tested by comparing the automated segmentation boundaries with those of the manual segmentation. For test purposes, five images from different regions of a follicular lymphoma slide were analyzed both by the proposed computerized method and human experts. The accuracy was evaluated with two measures of mean and their variations, including the true positive and false positive rates.
Results:
In our tests, the total number of follicular centers was 57. The resulting true positive and false positive rates were 74.40%?13.48% and 14.20%?5.28%, respectively.
Conclusion:
The preliminary experimental results indicate that the developed method is feasible as an initial step of the FL grading system, and that the approach exhibits a promising way that can be potentially applied to automatically detect image regions associated with follicle centers where the number of centroblasts can be automatically identified by further analysis.
