An automatic approach for heart segmentation in CT scans through image processing techniques and Concat-U-Net

Organs at risk (OARs) are healthy tissues around cancers that must be preserved in radiotherapy (RT). The heart is one of the fundamental organs for the full functioning of the human body. Protecting this organ in the RT is of paramount importance. For this, the planning process must be careful. Planning begins with manual segmentation by specialists in computed tomography (CT). However, manual segmentation combined with fatigue due to the number of slices to segment can cause human errors. Computational software has been developed for automatic heart segmentation in planning CT to assist specialists in this task. In this work, we propose an automatic deep learning method for heart segmentation from planning CT. The proposed method consists of 4 steps: (1) database acquisition from a public and diversified database; (2) volume standardization using registration and histogram matching; (3) coarse segmentation using atlas-based segmentation and a U-Net with a Concatenation Block (Concat-U-Net); and (4) fine segmentation using image processing techniques. We use a public database with 36 CTs who will undergo RT. This database is acquired from three different institutes. We achieve 95.25% of the Dice similarity coefficient, 87.95% of the Jaccard (JAC) Index, 96.71% of the sensitivity, and 99.39% of the accuracy. With the innovation of the proposed method and the promising results, we show that our method effectively uses heart segmentation. This method can serve as an ally to specialists and, with their expertise, can quickly treat patients undergoing RT treatments. •This work investigates a method for heart segmentation in planning CT.•It composed of atlas, CNN and image processing techniques.•The method was applied in 36 CT scans with an average of 200 slices.•We proposed a deep convolutional neural network with concatenation blocks.•The method achieved 95.25% of the Dice and 87.95% of Jaccard.