High-Throughput Prostate Cancer Gland Segmentation and Classification from Digitized Needle Core Biopsies.
Jun Xu, Rachel Sparks, Andrew Janowcyzk, John E Tomaszewski, Anant Madabhushi. Rutgers University, Piscataway, NJ; The Hospital of the University of Pennsylvania, Philadelphia; IIT Bombay, Piscataway, NJ
Background: CaP grading of histopathology can be divided into two separate tasks: identification of malignant regions and the Gleason grading of the malignant regions. Since manual discrimination of benign and malignant regions by pathologists is a time-consuming process, developing computerized decision support (CDS) systems that can quickly and accurately identify suspicious regions in tissue samples will enable the pathologist to focus their grading efforts on candidate regions, minimizing the time spent on identifying CaP regions. We present a high-throughput system for segmentation and classification of glands in high resolution digitized images, which will allow for rapid and accurate identification of suspicious regions on these samples.
Design: The high-throughput system includes the following three modules: 1) a hierarchical frequency weighted mean shift normalized cut for initial detection of glands; 2) a geodesic active contour model for gland segmentation; and 3) a diffeomorphic based similarity feature extraction and support vector machine for classification of glands as benign or cancerous.
Results: Classification accuracy in distinguishing benign from malignant glands over 23 H & E stained prostate studies when using the automated segmentation scheme was 82.5±9.1%, while the corresponding accuracy with manual segmentation was 82.89±3.97%; no statistically significant differences were identified between the two segmentation schemes. Figure 1 illustrates an example of whole-slide needle core biopsy of the prostate with malignant region delineated in blue. Glands labeled as benign/normal (black) and malignant (green) by the classifier are displayed.
Conclusions: We presented a high-throughput system for rapid and accurate gland segmentation and classification on high resolution digitized whole-slide images of needle core biopsy samples of the prostate. Since our system is able to automatically identify malignant regions with our CDS system, pathologists can save more time on analyzing the malignant regions.
Wednesday, March 2, 2011 9:30 AM
Poster Session V # 207, Wednesday Morning