Data augmentation based on spatial deformations for histopathology: An evaluation in the context of glomeruli segmentation.

BACKGROUND AND OBJECTIVE: The effective application of deep learning to digital histopathology is hampered by the shortage of high-quality annotated images. In this paper we focus on the supervised segmentation of glomerular structures in patches of whole slide images of renal histopathological slides. Considering a U-Net model employed for segmentation, our goal is to evaluate the impact of augmenting training data with random spatial deformations. METHODS: The effective application of deep learning to digital histopathology is hampered by the shortage of high-quality annotated images. In this paper we focus on the supervised segmentation of glomerular structures in patches of whole slide images of renal histopathological slides. Considering a U-Net model employed for segmentation, our goal is to evaluate the impact of augmenting training data with random spatial deformations. RESULTS: We show that augmenting training data with spatially deformed images yields an improvement of up to 0.23 in average Dice score, with respect to training with no augmentation. We demonstrate that deformations with relatively strong distortions yield the best performance increase, while previous work only report the use of deformations with low distortions. The selected deformation models yield similar performance increase, provided that their parameters are properly adjusted. We provide bounds on the optimal parameter values, obtained through parameter sampling, which is achieved in a lower computational complexity with our single-parameter method. The paper is accompanied by a framework for evaluating the impact of random spatial deformations on the performance of any U-Net segmentation model. CONCLUSION: To our knowledge, this study is the first to evaluate the impact of random spatial deformations on the segmentation of histopathological images. Our study and framework provide tools to help practitioners and researchers to make a better usage of random spatial deformations when training deep models for segmentation.

Pre-integrated volume rendering with non-linear gradient interpolation.

Shading is an important feature for the comprehension of volume datasets, but is difficult to implement accurately. Current techniques based on pre-integrated direct volume rendering approximate the volume rendering integral by ignoring non-linear gradient variations between front and back samples, which might result in cumulated shading errors when gradient variations are important and / or when the illumination function features high frequencies. In this paper, we explore a simple approach for pre-integrated volume rendering with non-linear gradient interpolation between front and back samples. We consider that the gradient smoothly varies along a quadratic curve instead of a segment in-between consecutive samples. This not only allows us to compute more accurate shaded pre-integrated look-up tables, but also allows us to more efficiently process shading amplifying effects, based on gradient filtering. An interesting property is that the pre-integration tables we use remain two-dimensional as for usual pre-integrated classification. We conduct experiments using a full hardware approach with the Blinn-Phong illumination model as well as with a non-photorealistic illumination model.

Per-pixel opacity modulation for feature enhancement in volume rendering.

Classical direct volume rendering techniques accumulate color and opacity contributions using the standard volume rendering equation approximated by alpha blending. However, such standard rendering techniques, often also aiming at visual realism, are not always adequate for efficient data exploration, especially when large opaque areas are present in a data set, since such areas can occlude important features and make them invisible. On the other hand, the use of highly transparent transfer functions allows viewing all the features at once, but often makes these features barely visible. In order to enhance feature visibility, we present in this paper a straightforward rendering technique that consists of modifying the traditional volume rendering equation. Our approach does not require an opacity transfer function, and instead is based on a function quantifying the relative importance of each voxel in the final rendering called relevance function. This function is subsequently used to dynamically adjust the opacity of the contributions per pixel. We conduct experiments with a number of possible relevance functions in order to show the influence of this parameter. As will be shown by our comparative study, our rendering method is much more suitable than standard volume rendering for interactive data exploration at a low extra cost. Thereby, our method avoids feature visibility restrictions without relying on a transfer function and yet maintains a visual similarity with standard volume rendering.