Medical Image Segmentation Combining Level Set Method and Deep Belief Networks

Abstract

Medical image segmentation is an important step in medical image analysis, where the main goal is the precise delineation of organs and tumours from medical images. For instance there is evidence in the field that shows a positive correlation between the precision of these segmentations and the accuracy observed in classification systems that use these segmentations as their inputs. Over the last decades, a vast number of medical image segmentation models have been introduced, where these models can be divided into five main groups: 1) image-based approaches, 2) active contour methods, 3) machine learning techniques, 4) atlas-guided segmentation and registration and 5) hybrid models. Image-based approaches use only intensity value or texture for segmenting (i.e., thresholding technique) and they usually do not produce precise segmentation. Active contour methods can use an explicit representation (i.e., snakes) with the goal of minimizing an energy function that forces the contour to move towards strong edges and maintains the contour smoothness. The use of implicit representation in active contour methods (i.e., level set method) embeds the contour as zero level set of a higher dimensional surface (i.e., the curve representing the contour does not need to be parameterized as in the Snakes model). Although successful, the main issue with active contour methods is the fact that the energy function must contain terms describing all possible shape and appearance variations, which is a complicated task given that it is hard to design by hand all these terms. Also, this type of active contour methods may get stuck at image regions that do not belong to the object of interest. Machine learning techniques address this issue by automatically learning shape and appearance models using annotated training images. Nevertheless, in order to meet the high accuracy requirements of medical image analysis applications, machine learning methods usually need large and rich training sets and also face the complexity of the inference process. Atlas-guided segmentation and registration use an atlas image, which is constructed based on manually segmentation images. The new image is segmented by registering it with the atlas image. These techniques have been applied successfully in many applications, but they still face some issues, such as their ability to represent the variability of anatomical structure and scale in medical image, and the complexity of the registration algorithms. In this work, we propose a new hybrid segmentation approach by combining a level set method with a machine learning approach (deep belief network). Our main objective with this approach is to achieve segmentation accuracy results that are either comparable or better than the ones produced with machine learning methods, but using relatively smaller training sets. These weaker requirements on the size of training sets is compensated by the hand designed segmentation terms present in typical level set methods, that are used as prior information on the anatomy to be segmented (e.g., smooth contours, strong edges, etc.). In addition, we choose a machine learning methodology that typically requires smaller annotated training sets, compared to other methods proposed in this field. Specifically, we use deep belief networks, with training sets consisting to a large extent of un-annotated training images. In general, our hybrid segmentation approach uses the result produced by the deep belief network as a prior in the level set evolution. We validate this method on the Medical Image Computing and Computer Assisted Intervention (MICCAI) 2009 left ventricle segmentation challenge database and on the Japanese Society of Radiological Technology (JSRT) lung segmentation dataset. The experiments show that our approach produces competitive results in the field in terms of segmentation accuracy. More specifically, we show that the use of our proposed methodology in a semi-automated segmentation system (i.e., using a manual initialization) produces the best result in the field in both databases above, and in the case of a fully automated system, our method shows results competitive with the current state of the art.