Utilizing Image Segmentation and Supervised Machine Learning Algorithms to Detect Prostatic Carcinoma in Histological Images
Abstract
Prostate cancer (PCa) exists as one of the most prevalent forms of cancer in men. It has been found that approximately 23% of men who test negative for PCa during their first needle biopsy are diagnosed with PCa during subsequent biopsies. Furthermore, it has been observed that existing prostate lesions are not identified in 2.7% of prostate needle biopsies. A need exists, then, to obtain a more accurate and consistent means of detecting the presence of prostate cancer in histological images. The purpose of this project was to determine if an algorithmic approach can be used to increase the accuracy of detecting prostatic lesions. Initial efforts were focused on the detection of patterns one, two, and three of acinar adenocarcinoma by isolating gland lumen via image segmentation. Results show promise in isolating malignant glands according to lumen size. Further work will be done to qualify image-segmented lumen as malignant via k-Nearest Neighbors classification of bordering nuclei according to size and color. Current progress suggests potential for accurate algorithmic detection of prostate cancer.
Utilizing Image Segmentation and Supervised Machine Learning Algorithms to Detect Prostatic Carcinoma in Histological Images
Prostate cancer (PCa) exists as one of the most prevalent forms of cancer in men. It has been found that approximately 23% of men who test negative for PCa during their first needle biopsy are diagnosed with PCa during subsequent biopsies. Furthermore, it has been observed that existing prostate lesions are not identified in 2.7% of prostate needle biopsies. A need exists, then, to obtain a more accurate and consistent means of detecting the presence of prostate cancer in histological images. The purpose of this project was to determine if an algorithmic approach can be used to increase the accuracy of detecting prostatic lesions. Initial efforts were focused on the detection of patterns one, two, and three of acinar adenocarcinoma by isolating gland lumen via image segmentation. Results show promise in isolating malignant glands according to lumen size. Further work will be done to qualify image-segmented lumen as malignant via k-Nearest Neighbors classification of bordering nuclei according to size and color. Current progress suggests potential for accurate algorithmic detection of prostate cancer.