Fully-Automated Segmentation of Tumor Areas in Tissue Confocal Images - Comparison between a Custom Unsupervised and a Supervised SVM Approach

In this paper we present a fully-automated method for the detection of tumor areas in immunohistochemical confocal images. The image segmentation provided by the proposed technique allows quantitative protein activity evaluation on the target tumoral tissue disregarding tissue areas that are not affected by the pathology, such as connective tissue. The automated method, that is based on an innovative unsupervised clustering approach, enables more accurate tissue segmentation compared to traditional supervised methods that can be found in literature, such as Support Vector Machine (SVM). Experimental results conducted on a large set of heterogeneous immunohistochemical lung cancer tissue images demonstrate that the proposed approach overcomes the performance of SVM by 8%, achieving on average an accuracy of 90%.