Unconstrained environments with variable ambient illumination and changes of head pose are still challenging for many face recognition systems. To recognize a person independent of pose, we first fit an active appearance model to a given facial image. Shape information is used to transform the face into a pose-normalized representation. We decompose the transformed face into local regions and extract texture features from these not necessarily rectangular regions using a shape-adapted discrete cosine transform. We show that these features contain sufficient discriminative information to recognize persons across changes in pose. Furthermore, our experimental results show a significant improvement in face recognition performance on faces with pose variations when compared with a block-DCT based feature extraction technique in an access control scenario.
Poster: ECR 2020 / C-05625 / AI-based positioning quality check for chest x-ray by: J. von Berg, D. Bystrov, A. Goosen, S. Kronke, M. Bruck, T. Harder, N. Wieberneit , S. Young; Hamburg/DE
In medical X-ray examinations, images suffer considerably from severe, signal-dependent noise as a result of the effort to keep applied doses as low as possible. This noise can be seen as an additive signal that degrades image quality and might disguise valuable content. Lost information has to be restored in a post-processing step. The crucial aspect of filtering medical images is preservation of edges and texture on the one hand and removing noise on the other hand. Classical smoothing filters, such as Gaussian or box filtering. are data-independent and equally blur the image content. State-of-the-art methods currently make use of local neighborhoods or global image statistics. However, exploiting global self-similarity within an image and inter-image similarity for subsequent frames of a sequence bears an unused potential for image restoration. We introduce a non-local filter with data-dependent response that closes the gap between local filtering and stochastic methods. The filter is based on the non-local means approach proposed by Buades1 et al. and is similar to bilateral filtering. In order to apply this approach to medical data, we heavily reduce the computational costs incurred by the original approach. Thus it is possible to interactively enhance single frames or selected regions of interest within a sequence. The proposed filter is applicable for time-domain filtering without the need for accurate motion estimation. Hence it can be seen as a general solution for filtering 2D as well as 2D+t X-ray image data.
Breast density has become an established risk indicator for developing breast cancer. Current clinical practice reflects this by grading mammograms patient-wise as entirely fat, scattered fibroglandular, heterogeneously dense, or extremely dense based on visual perception. Existing (semi-) automated methods work on a per-image basis and mimic clinical practice by calculating an area fraction of fibroglandular tissue (mammographic percent density). We suggest a method that follows clinical practice more strictly by segmenting the fibroglandular tissue portion directly from the joint data of all four available mammographic views (cranio-caudal and medio-lateral oblique, left and right), and by subsequently calculating a consistently patient-based mammographic percent density estimate. In particular, each mammographic view is first processed separately to determine a region of interest (ROI) for segmentation into fibroglandular and adipose tissue. ROI determination includes breast outline detection via edge-based methods, peripheral tissue suppression via geometric breast height modeling, and - for medio-lateral oblique views only - pectoral muscle outline detection based on optimizing a three-parameter analytic curve with respect to local appearance. Intensity harmonization based on separately acquired calibration data is performed with respect to compression height and tube voltage to facilitate joint segmentation of available mammographic views. A Gaussian mixture model (GMM) on the joint histogram data with a posteriori calibration guided plausibility correction is finally employed for tissue separation. The proposed method was tested on patient data from 82 subjects. Results show excellent correlation (r = 0.86) to radiologist's grading with deviations ranging between -28%, (q = 0.025) and +16%, (q = 0.975).
The automatic detection of critical findings in chest X-rays (CXR), such as pneumothorax, is important for assisting radiologists in their clinical workflow like triaging time-sensitive cases and screening for incidental findings. While deep learning (DL) models has become a promising predictive technology with near-human accuracy, they commonly suffer from a lack of explainability, which is an important aspect for clinical deployment of DL models in the highly regulated healthcare industry. For example, localizing critical findings in an image is useful for explaining the predictions of DL classification algorithms. While there have been a host of joint classification and localization methods for computer vision, the state-of-the-art DL models require locally annotated training data in the form of pixel level labels or bounding box coordinates. In the medical domain, this requires an expensive amount of manual annotation by medical experts for each critical finding. This requirement becomes a major barrier for training models that can rapidly scale to various findings. In this work, we address these shortcomings with an interpretable DL algorithm based on multi-instance learning that jointly classifies and localizes critical findings in CXR without the need for local annotations. We show competitive classification results on three different critical findings (pneumothorax, pneumonia, and pulmonary edema) from three different CXR datasets.
