Automatic Segmentation Of Lung Carcinoma In Histological Images Using A Visual Dictionary

The quantification of remaining vital tumor tissue in patients undergoing chemo-/radiotherapy is necessary to assess the response to treatment. In this work, we present an automatic segmentation of pan-cytokeratin stained histological images of non-small cell lung carcinoma, which provides the ratio of vital vs. necrotic tumor tissue. The proposed method learns from training patches of vital and necrotic tissues, from which a set of features are extracted. Image superpixels are then labeled according to their nearest neighbors in the training feature space. Segmentation results were quantitatively assessed using leave-one-out tests. The proposed method achieved average sensitivity (specificity) of 91% (95%) and 73% (96%) for vital tumor epithelia and necrosis, respectively.