Computer Vision News - June 2019

16 Computer Vision News Challenges in Lung Tumor Segmentation Every month, Computer Vision News reviews a successful project. Our main purpose is to show how diverse image processing techniques contribute to solving technical challenges and real world constraints. This month we review challenges and solutions in a Medical Imaging project by RSIP Vision : Challenges in Lung Tumor Segmentation and Classification . Project Computerized Tomography (CT) of the lungs is a common procedure to assess existence and character of pulmonary lesions , including nodules and masses: Although some cases are benign, it is crucial to detect the cancerous ones correctly in order to achieve clear and early diagnosis. The order of nodule classification is the following: first, according to their transparency density - some of them are solid and occlude all tissues behind them - others are almost transparent and their look similar to hazy glass is called Ground- Glass Opacity (GGO) . In between we find also semi-solid nodules which combine these two features: these have the highest chances to be malignant tumors . Second, according to the smoothness of their borders - the most problematic nodules being those with spiculated borders that have protruding spikes like sun rays. Nodules differ very much one from another, both in shape and density ; in some cases they might have non-solid/ ground-glass appearance. The classification of the nodules into solid, semi-solid or non solid appearance aids the physician in assessing the probability of malignancy . It is not only the appearance of the nodule that counts, but also the shape of its borders . Smooth borders increase the chance of it being benign, as apposed to irregular, spiculated or lobulated borders, that increase the chance of malignancy. Volume also is key, as larger nodules and masses are more suspect than smaller ones. Besides their large diversity, the algorithm developer encounters another challenge at the location of the lesion: they can be found anywhere within the lung and even touching the hilum or mediastinum, making it difficult even to expert radiologists to recognize where the boundaries are and whether the nodule belongs to the airway system or not. Other findings may appear like pulmonary nodules while they are not. This might be tricky, making room for false positive findings. Possible misleading findings include blood vessels, atelectasis or fluid accumulation or scar tissue. We facilitated the algorithm by annotating with a different color the accumulations of fluid inside the lungs. While there are problematic spots, they are not nodules and therefore they do not interest our study. We also classified the different appearances of lesions by color - red is solid, green is combined, blue is GGO . This was one of the main tricks that we used to power the algorithm and minimize the false positives rate . “This was one of the main tricks that we used” by Michal Margalit