Matthias Dorfer

Creating a Large-Scale Silver Corpus from Multiple Algorithmic Segmentations

Currently, increasingly large medical imaging data sets become available for research and are analysed by a range of algorithms segmenting anatomical structures automatically and interactively. While they provide segmentations on a much larger scale than possible to achieve with expert annotators, they are typically less accurate than experts. We present and compare approaches to estimate segmentations on large imaging data sets based on a small number of expert annotated examples, and algorithmic segmentations on a much larger data set. Results demonstrate that combining algorithmic segmentations is reliably outperforming the average individual algorithm. Furthermore, injecting organ specific reliability assessments of algorithms based on expert annotations improves accuracy compared to standard label fusion algorithms. The proposed methods are particularly relevant in putting the results of large image analysis algorithm benchmarks to long-term use.

Music Playlist Continuation by Learning from Hand-Curated Examples and Song Features: Alleviating the Cold-Start Problem for Rare and Out-of-Set Songs

Automated music playlist generation is a specific form of music recommendation. Generally stated, the user receives a set of song suggestions defining a coherent listening session. We hypothesize that the best way to convey such playlist coherence to new recommendations is by learning it from actual curated examples, in contrast to imposing ad hoc constraints. Collaborative filtering methods can be used to capture underlying patterns in hand-curated playlists. However, the scarcity of thoroughly curated playlists and the bias towards popular songs result in the vast majority of songs occurring in very few playlists and thus being poorly recommended. To overcome this issue, we propose an alternative model based on a song-to-playlist classifier, which learns the underlying structure from actual playlists while leveraging song features derived from audio, social tags and independent listening logs. Experiments on two datasets of hand-curated playlists show competitive performance compared to collaborative filtering when sufficient training data is available and more robust performance when recommending rare and out-of-set songs. For example, both approaches achieve a recall@100 of roughly 35% for songs occurring in 5 or more training playists, whereas the proposed model achieves a recall@100 of roughly 15% for songs occurring in 4 or less training playlists, compared to the 3% achieved by collaborative filtering.

Drum transcription from polyphonic music with recurrent neural networks

Automatic drum transcription methods aim at extracting a symbolic representation of notes played by a drum kit in audio recordings. For automatic music analysis, this task is of particular interest as such a transcript can be used to extract high level information about the piece, e.g., tempo, downbeat positions, meter, and genre cues. In this work, an approach to transcribe drums from polyphonic audio signals based on a recurrent neural network is presented. Deep learning techniques like dropout and data augmentation are applied to improve the generalization capabilities of the system. The method is evaluated using established reference datasets consisting of solo drum tracks as well as drums mixed with accompaniment. The results are compared to state-of-the-art approaches on the same datasets. The evaluation reveals that F-measure values higher than state of the art can be achieved using the proposed method.

Robust, automatic real-time monitoring of the time course of the individual alpha frequency in the time and frequency domain

We analyzed three different approaches to automatic real-time monitoring of the time course of individual alpha frequencies (IAFs) of the human electro-encephalograms. Fast Fourier transform and wavelet transform were compared to classical automated cycle counting in the time domain. With fast Fourier and wavelet transform, test results with healthy adult subjects, demented and psychiatric patients revealed typical short-term variations of the instantaneous IAFs of about ± 2 Hz. When cycles were counted in the time domain, however, variations of only ± 1 Hz were recorded. Thus, IAF measurement in the time domain appears to be particularly suitable. We also observed long-term IAF trends that typically amounted to about ± 0.5 to ± 1.0 Hz. Therefore, our hypothesis is that the IAF does not constitute an intra-individual constant but varies with time and cognitive state. Our fully automatic real-time signal-processing procedure includes pre-processing for artifact detection and for localization of segments with synchronized alpha oscillations where the IAF should preferably be measured.

End-to-end cross-modality retrieval with CCA projections and pairwise ranking loss

Cross-modality retrieval encompasses retrieval tasks where the fetched items are of a different type than the search query, e.g., retrieving pictures relevant to a given text query. The state-of-the-art approach to cross-modality retrieval relies on learning a joint embedding space of the two modalities, where items from either modality are retrieved using nearest-neighbor search. In this work, we introduce a neural network layer based on canonical correlation analysis (CCA) that learns better embedding spaces by analytically computing projections that maximize correlation. In contrast to previous approaches, the CCA layer allows us to combine existing objectives for embedding space learning, such as pairwise ranking losses, with the optimal projections of CCA. We show the effectiveness of our approach for cross-modality retrieval on three different scenarios (text-to-image, audio-sheet-music and zero-shot retrieval), surpassing both Deep CCA and a multi-view network using freely learned projections optimized by a pairwise ranking loss, especially when little training data is available (the code for all three methods is released at: https://github.com/CPJKU/cca_layer).

Associating approximate paths and temporal sequences of noisy detections: Application to the recovery of spatio-temporal cancer cell trajectories.

In this paper we address the problem of recovering spatio-temporal trajectories of cancer cells in phase contrast video-microscopy where the user provides the paths on which the cells are moving. The paths are purely spatial, without temporal information. To recover the temporal information associated to a given path we propose an approach based on automatic cell detection and on a graph-based shortest path search. The nodes in the graph consist of the projections of the cell detections onto the geometrical cell path. The edges relate nodes which correspond to different frames of the sequence and potentially to the same cell and trajectory. In this directed graph we search for the shortest path and use it to define a temporal parametrization of the corresponding geometrical cell path. An evaluation based on 286 paths of 7 phase contrast microscopy videos shows that our algorithm allows to recover 92% of trajectory points with respect to the associated ground truth. We compare our method with a state-of-the-art algorithm for semi-automated cell tracking in phase contrast microscopy which requires interactively placed starting points for the cells to track. The comparison shows that supporting geometrical paths in combination with our algorithm allow us to obtain more reliable cell trajectories.

Constructing an Un-biased Whole Body Atlas from Clinical Imaging Data by Fragment Bundling

Atlases have a tremendous impact on the study of anatomy and function, such as in neuroimaging, or cardiac analysis. They provide a means to compare corresponding measurements across populations, or model the variability in a population. Current approaches to construct atlases rely on examples that show the same anatomical structure (e.g., the brain). If we study large heterogeneous clinical populations to capture subtle characteristics of diseases, we cannot assume consistent image acquisition any more. Instead we have to build atlases from imaging data that show only parts of the overall anatomical structure. In this paper we propose a method for the automatic contruction of an un-biased whole body atlas from so-called fragments. Experimental results indicate that the fragment based atlas improves the representation accuracy of the atlas over an initial whole body template initialization.

Recursive water flow: A shape decomposition approach for cell clump splitting

We propose a hierarchical method for splitting cell clumps into individual cells which we call Recursive Water Flow (rwf). For the segmentation of cells in histological images we first apply foreground segmentation leading to connected regions of clumping cells. rwf defines one-dimensional cost-functions along paths on the skeleton of these regions. In particular, we collect split-relevant information along geodesic paths between skeleton end-points. This framework allows to exploit combined shape and intensity information for identifying optimal split positions and recursively decomposing the clumps into single distinct cells. Results show that our framework helps to improve cell segmentation performance in contrast to splitting based on local geometrical clues.