Medical imaging

Analysis of medical images through visual inspection presents several difficulties, due to the subjectivity of the physician, the complexity of the data, the variety of image acquisition modalities and the difficulty for visual analysis of volumetric data, and even volumetric temporal sequences. The application of signal processing techniques in the medical imaging field has been producing advances in diagnosis assistance during the latest decades. Some of its applications include enhancement and detection of target patterns, segmentation and parametrization of tissues, and multimodal registration. Moreover, in combination with current 3D visualization techniques, it eases visual inspection and it even allows the introduction of augmented reality systems for surgical planning.

One of the applications that has received the most attention in the latest years is the segmentation of organs or pathological tissues. Computational segmentation methods not only facilitate the demarcation of tissues from noisy and complex data, but it also allows rapid, automatic and repeatable extraction of geometric parameters (diameter, area, volume, shape, …) of great help for patient diagnosis and monitoring.

Among the multiple existing segmentation methods, deformable models have stand out due to their combined capability for regularization, by means of local shape constraints, and incorporation of prior knowledge. They are dynamic models that, starting from an initial state, are able to evolve until they fit the low level image features, with ability to manage topological changes, as observed in the following image. Additionally, they are suitable for tracking of deformable objects in video sequences.

Gradiant has personnel with experience in digital medical image processing and analysis, both from 2D and 3D data (CT, MRI, SPECT, …), and in particular, in 3D segmentation with deformable models. We have worked with geodesic models and model initialization using multiresolution features. Our background in the field of pattern recognition can be applied to the most recent segmentation techniques, which combine deformable models with region classification methods based on appearance statistical models.

Besides, our experience in multi- and hyper-spectral image acquisition and analysis is also of application to medical image segmentation, especially in dermatology, where this image modality is being successfully used for detection and monitoring of melanoma. The analysis of the information provided by the different frequency bands allows revealing pathological areas that are not observable in the visible range, like the area tagged in yellow in the image.