A Content-Based Image Retrieval System for Digitized Prostate Histopathology.
Akshay Sridhar, Scott Doyle, Anant Madabhushi. Rutgers University, Piscataway, NJ
Background: In a content-based image retrieval system (CBIR) system, the user inputs a query image and the system outputs images from the database that are most similar to the query image. Such a system in the context of prostate cancer pathology would be an invaluable educational, research, and clinical tool for pathologists. For instance, approximately 1 million prostate cancer needle core biopsies, with a corresponding 15 million prostate tissue samples, are generated annually. However, only 1 million samples are positive for cancer. The CBIR companion tool can help reduce the time pathologists spend looking at benign samples, allowing them to focus on the suspicious samples. Our new CBIR system utilizes a weighted distance metric (WDM) and we apply it to the problem of distinguishing benign from malignant prostate cancer biopsy images.
Design: H&E stained prostate histopathology images were obtained from 58 patients and 14 texture features were extracted from each image. Each texture feature in conjunction with a Bayesian classifier was used to build a weak learner. The machine learning algorithm Adaboost was then used in conjunction with these weak learners to learn feature weights which resulted in maximum discriminability between the benign and malignant classes in a training set. The weights obtained were used in conjunction with the Euclidian distance metric resulting in a WDM. The WDM is then used to build a CBIR system to retrieve database samples that are most similar to the query image.
Results: Precision-Recall (PR) curves were used to evaluate the WDM-CBIR and compared against a CBIR system that used the Euclidian distance metric (EDM-CBIR) with the same texture features. Figure 1a clearly shows that our WDM-CBIR outperformed the EDM-CBIR on a cohort of 58 prostate cancer studies. The average area under the curve for the WDM-CBIR was 0.95 compared to 0.72 for the EDM-CBIR.
Conclusions: We introduced a new WDM-CBIR system that differentially weights individual image attributes based on their discriminatory performance. WDM-CBIR outperformed the EDM-CBIR on 58 prostate cancer studies. WDM-CBIR can be extended and applied to other domains within digital pathology.
Tuesday, March 1, 2011 1:00 PM
Poster Session IV # 187, Tuesday Afternoon