An Interactive Statistical Image Segmentation and Visualization System
Supervised learning can be used to segment / identify regions of interest in images making use of color and morphological information. A novel object identiﬁcation algorithm was developed in Java to locate immune and cancer cells in images of immunohistochemically-stained lymph node tissue from the recent Kohrt study and also shows promise in other domains. Our method leans heavily on the use of color and the relative homogeneity of object appearance. As is often the case in segmentation, an algorithm speciﬁcally tailored to the application works better than using broader methods that work passably well on any problem. Our main innovation is interactive feature extraction from color images. We also enable the user to improve the classiﬁcation with an interactive visualization system. This is then coupled with the statistical learning algorithms and intensive interaction with the user over many classiﬁcation-correction iterations, resulting in a highly accurate and user-friendly solution. The system ultimately provides the locations of every cell in the entire tissue. This data can be analyzed using standard statistical methods such as spatial analyses or robust clustering. This data is invaluable in the study of multidimensional relationships between cell populations and tumor structure.
We start with an overview of current practices in image recognition and a short presentation of the clinical context that motivated this research, we then describe the software and the complete work- ﬂow involved, ﬁnally the last two sections present technical details and potential improvements. Our interactive algorithm, although developed to solve a speciﬁc problem in histology, works on a wide variety of images. For instance, locating of oranges in a photograph of an orange grove (see Fig.1). Any image that has few relevant colors, such as green and orange in the above example, where the objects of interest vary little in shape, size, and color, can be analyzed using our algorithm. First, we will describe the application to cell recognition in microscopic images.
As emphasized in recent reviews automated computer vision techniques applied to microscopy images are transforming the ﬁeld of pathology. Computerized vision techniques automate cell type recognition Funded by a DOD Era Hope Scholar grant to PPL and by NSF-DMSand they enable a more objective approach to cell classiﬁcation. These techniques provide at the same time a hierarchy of quantitative features measured on the images. Recent work on character recognition shows how efﬁcient interactivity can be in image recognition problems, with the user pointing out mistakes in real time, thus providing online improvement. In modern jargon, we call this interactive boosting . Current cell image analysis systems such as EBimage and Midas do not provide these interactive visualization and correction features. 1.2 Speciﬁc context: Breast Cancer Prognosis Kohrt et al (2005) showed that breast cancer prognosis could be greatly improved by using immune population information from immunohistochemically-stained lymph nodes. To take his analysis a step further, we would like to detect and pinpoint the location of each and every cancer and immune cell in the high-resolution full-mount images of lymph nodes acquired via automated microscopy. This task is harder than classiﬁcation of an entire slide as normal or abnormal as done in . A typical tissue contains a variety of regions characteristic of cancer, immune cells, or both. It would not be possible for a histopathologist to identify and count all the cells of each type on such a slide. Even if this was feasible employing a team, the problems of subjectivity and bias on such a scale will discredit the results. It would be very useful to have an automated system to identify cells.