Saliency methods are critical tools that allow the estimation of the most important features of an input image that contribute to the network’s prediction. These tools are pivotal in high-stakes applications such as medical diagnosis or autonomous driving. Additionally, these tools can help identify models’ biasedness, such as a strong prior on object placement, easily distinguishable background features, or frequent object co-occurrence. We introduce RankPix, a novel saliency method for visual bias identification in image classification tasks. RankPix is a derivative-free approach that allows the identification of a minimum subset of pixels/features at a given network layer that changes the output of a classifier. Surprisingly, this approaches provides equivalent performance to gradient-based approaches on the standard pointing game benchmark. More interestingly, RankPix outperforms traditional approaches for systematic bias identification.

There are strong incentives to build classification systems that show outstanding performance on various datasets and benchmarks. This can encourage a narrow focus on models and the performance metrics used to evaluate and compare them—resulting in a growing body of literature to evaluate and compare metrics. This paper strives for a more balanced perspective on binary classifier performance metrics by showing how uncertainty in these metrics can easily eclipse differences in empirical performance. We emphasise the discrete nature of confusion matrices and show how they can be well represented in a 3D lattice whose cross-sections form the space of receiver operating characteristic (ROC) curves. We develop novel interactive visualisations of performance metric contours within (and beyond) ROC space, showing the discrete probability mass functions of true and false positive rates and how these relate to performance metric distributions. We aim to raise awareness of the substantial uncertainty in performance metric estimates that can arise when classifiers are evaluated on empirical datasets and benchmarks, and that performance claims should be tempered by this understanding.

Cortical thickness (CTh) is routinely used to quantify grey matter atrophy as it is a significant biomarker in studying neurodegenerative and neurological conditions. Clinical studies commonly employ one of several available CTh estimation software tools to estimate CTh from brain MRI scans. In recent years, machine learning-based methods emerged as a faster alternative to the main-stream CTh estimation methods (e.g. FreeSurfer). Evaluation and comparison of CTh estimation methods often include various metrics and downstream tasks, but none fully covers the sensitivity to sub-voxel atrophy characteristic of neurodegeneration. In addition, current evaluation methods do not provide a framework for the intra-method region-wise evaluation of CTh estimation methods. Therefore, we propose a method for brain MRI synthesis capable of generating a range of sub-voxel atrophy levels (global and local) with quantifiable changes from the baseline scan. We further create a synthetic test set and evaluate four different CTh estimation methods: FreeSurfer (cross-sectional), FreeSurfer (longitudinal), DL+DiReCT and HerstonNet. DL+DiReCT showed superior sensitivity to sub-voxel atrophy over other methods in our testing framework. The obtained results indicate that our synthetic test set is suitable for benchmarking CTh estimation methods on both global and local scales as well as regional inter-and intra-method performance comparison.

Through the years, several CAD systems have been developed to help radiologists in the hard task of detecting signs of cancer in mammograms. In these CAD systems, mass segmentation plays a central role in the decision process. In the literature, mass segmentation has been typically evaluated in a intra-sensor scenario, where the methodology is designed and evaluated in similar data. However, in practice, acquisition systems and PACS from multiple vendors abound and current works fails to take into account the differences in mammogram data in the performance evaluation. In this work it is argued that a comprehensive assessment of the mass segmentation methods requires the design and evaluation in datasets with different properties. To provide a more realistic evaluation, this work proposes: a) improvements to a state of the art method based on tailored features and a graph model; b) a head-to-head comparison of the improved model with recently proposed methodologies based in deep learning and structured prediction on four reference databases, performing a cross-sensor evaluation. The results obtained support the assertion that the evaluation methods from the literature are optimistically biased when evaluated on data gathered from exactly the same sensor and/or acquisition protocol.

In this paper, we introduce and evaluate the systems submitted to the first Overlapping Cervical Cytology Image Segmentation Challenge, held in conjunction with the IEEE International Symposium on Biomedical Imaging 2014. This challenge was organized to encourage the development and benchmarking of techniques capable of segmenting individual cells from overlapping cellular clumps in cervical cytology images, which is a prerequisite for the development of the next generation of computer-aided diagnosis systems for cervical cancer. In particular, these automated systems must detect and accurately segment both the nucleus and cytoplasm of each cell, even when they are clumped together and, hence, partially occluded. However, this is an unsolved problem due to the poor contrast of cytoplasm boundaries, the large variation in size and shape of cells, and the presence of debris and the large degree of cellular overlap. The challenge initially utilized a database of 16 high-resolution ( ×40 magnification) images of complex cellular fields of view, in which the isolated real cells were used to construct a database of 945 cervical cytology images synthesized with a varying number of cells and degree of overlap, in order to provide full access of the segmentation ground truth. These synthetic images were used to provide a reliable and comprehensive framework for quantitative evaluation on this segmentation problem. Results from the submitted methods demonstrate that all the methods are effective in the segmentation of clumps containing at most three cells, with overlap coefficients up to 0.3. This highlights the intrinsic difficulty of this challenge and provides motivation for significant future improvement.

We describe an automated methodology for the analysis of unregistered cranio-caudal (CC) and medio-lateral oblique (MLO) mammography views in order to estimate the patient's risk of developing breast cancer. The main innovation behind this methodology lies in the use of deep learning models for the problem of jointly classifying unregistered mammogram views and respective segmentation maps of breast lesions (i.e., masses and micro-calcifications). This is a holistic methodology that can classify a whole mammographic exam, containing the CC and MLO views and the segmentation maps, as opposed to the classification of individual lesions, which is the dominant approach in the field. We also demonstrate that the proposed system is capable of using the segmentation maps generated by automated mass and micro-calcification detection systems, and still producing accurate results. The semi-automated approach (using manually defined mass and micro-calcification segmentation maps) is tested on two publicly available data sets (INbreast and DDSM), and results show that the volume under ROC surface (VUS) for a 3-class problem (normal tissue, benign, and malignant) is over 0.9, the area under ROC curve (AUC) for the 2-class "benign versus malignant" problem is over 0.9, and for the 2-class breast screening problem (malignancy versus normal/benign) is also over 0.9. For the fully automated approach, the VUS results on INbreast is over 0.7, and the AUC for the 2-class "benign versus malignant" problem is over 0.78, and the AUC for the 2-class breast screening is 0.86.

In this paper we investigate multiple instance learning (MIL), using generic tile-based spatio-temporal features, for the classification of benign and malignant lesions in breast cancer magnetic resonance imaging (MRI). In particular, we compare the performance of citation-kNN (CkNN) and conventional kNN against a traditional approach based on bespoke features extracted from a segmented region-of-interest (ROI). Results demonstrate that tile-based CkNN has equivalent performance to ROI-based classification. However, the tile-based approach does not require any domain specific features typically used in breast MRI. This not only has the potential to make tile-based classification robust to inaccuracies in the delineation of suspicious lesions, but also makes it suitable for the detection of suspicious lesions prior to segmentation.

We determine the power spectrum of an ideal neuron which encodes information using a pulse interval modulation scheme in continuous time. We develop this by considering the rigorous derivation of the Digital Pulse Interval Modulation (DPIM) coding scheme spectra of L. Vangelista et al. in the limit of the coding slot size approaching zero. We show in this limit the spectrum is identical to that of a filtered renewal process frequently used to model neuroscience time series data. Using this renewal theory equivalence we then use the `Fundamental Isometry Theorem' developed by Win & Ridolfi to show that introducing firing time jitter (as a simple model for noise effects) removes non-Poisson structure and reduces the utility of spectral feature selection. Lastly we show with sufficient jittering that the Bartlett spectrum of any renewal process reduces to that of a Poisson process, with a spectral density consistent with Carson's theorem for shot noise.

In this paper we demonstrate by simulation, that the spike features apparent in low-impedance deep brain stimulation (DBS) targeting microelectrode recordings (MER) may not reflect the action potentials of individual neurons. Rather, they are more likely to be compound action potentials from a synchronized group of neurons local to the electrode. Initially we simulate the MER by combining the electric fields from a large number of independent neurons surrounding the microelectrode tip. When synchronization is introduced amongst neurons the resulting discernible spikes in an MER are far more likely to relate to compound action potentials from subsets of synchronized neurons than individual action potentials. Different sub-sets of neurons are then synchronized to see how well a conventional spike sorting algorithm can differentiate the compound action potentials from different groups of neurons. These simulations offer insight into the clinical interpretation of DBS MERs used to target deep brain structures.

In this paper, we present a novel method for the segmentation of breast masses from mammograms exploring structured and deep learning. Specifically, using structured support vector machine (SSVM), we formulate a model that combines different types of potential functions, including one that classifies image regions using deep learning. Our main goal with this work is to show the accuracy and efficiency improvements that these relatively new techniques can provide for the segmentation of breast masses from mammograms. We also propose an easily reproducible quantitative analysis to assess the performance of breast mass segmentation methodologies based on widely accepted accuracy and running time measurements on public datasets, which will facilitate further comparisons for this segmentation problem. In particular, we use two publicly available datasets (DDSM-BCRP and INbreast) and propose the computation of the running time taken for the methodology to produce a mass segmentation given an input image and the use of the Dice index to quantitatively measure the segmentation accuracy. For both databases, we show that our proposed methodology produces competitive results in terms of accuracy and running time.

In this paper, we present an improved algorithm for the segmentation of cytoplasm and nuclei from clumps of overlapping cervical cells. This problem is notoriously difficult because of the degree of overlap among cells, the poor contrast of cell cytoplasm and the presence of mucus, blood, and inflammatory cells. Our methodology addresses these issues by utilizing a joint optimization of multiple level set functions, where each function represents a cell within a clump, that have both unary (intracell) and pairwise (intercell) constraints. The unary constraints are based on contour length, edge strength, and cell shape, while the pairwise constraint is computed based on the area of the overlapping regions. In this way, our methodology enables the analysis of nuclei and cytoplasm from both free-lying and overlapping cells. We provide a systematic evaluation of our methodology using a database of over 900 images generated by synthetically overlapping images of free-lying cervical cells, where the number of cells within a clump is varied from 2 to 10 and the overlap coefficient between pairs of cells from 0.1 to 0.5. This quantitative assessment demonstrates that our methodology can successfully segment clumps of up to 10 cells, provided the overlap between pairs of cells is <;0.2. Moreover, if the clump consists of three or fewer cells, then our methodology can successfully segment individual cells even when the overlap is ~0.5. We also evaluate our approach quantitatively and qualitatively on a set of 16 extended depth of field images, where we are able to segment a total of 645 cells, of which only ~10% are free-lying. Finally, we demonstrate that our method of cell nuclei segmentation is competitive when compared with the current state of the art.

Steady-State Visual Evoked Potential-based (SSVEP) Brain-Computer Interface (BCI) shows great potential as a viable BCI due to its ease of implementation and speed. However, the majority of the SSVEP-BCI implementations use only features from the Power Spectral Density (PSD) despite the fact that upon transforming the signals to the Fourier domain, both the phase and amplitude components are available. In this study we extract the phase response and compute the phase variance as a measure of phase synchrony. This phase synchrony method is called Component Synchrony Measure (CSM). Our results indicate that by including the CSM as a feature, the SSVEP-BCI classification accuracy is significantly enhanced. This further establishes the use of both amplitude and phase information for obtaining good classification accuracy in SSVEP-BCI.

Diabetes mellitus is a chronic disease and a major public health challenge worldwide. According to the International Diabetes Federation, there are currently 246 million diabetic people worldwide, and this number is expected to rise to 380 million by 2025. Furthermore, 3.8 million deaths are attributable to diabetes complications each year. It has been shown that 80% of type 2 diabetes complications can be prevented or delayed by early identification of people at risk. In this context, several data mining and machine learning methods have been used for the diagnosis, prognosis, and management of diabetes. In this paper, we propose utilizing support vector machines (SVMs) for the diagnosis of diabetes. In particular, we use an additional explanation module, which turns the "black box" model of an SVM into an intelligible representation of the SVM's diagnostic (classification) decision. Results on a real-life diabetes dataset show that intelligible SVMs provide a promising tool for the prediction of diabetes, where a comprehensible ruleset have been generated, with prediction accuracy of 94%, sensitivity of 93%, and specificity of 94%. Furthermore, the extracted rules are medically sound and agree with the outcome of relevant medical studies.

This paper presents a new algorithm for denoising dynamic contrast-enhanced (DCE) MR images. It is a novel variation on the nonlocal means (NLM) algorithm. The algorithm, called dynamic nonlocal means (DNLM), exploits the redundancy of information in the temporal sequence of images. Empirical evaluations of the performance of the DNLM algorithm relative to seven other denoising methods-simple Gaussian filtering, the original NLM algorithm, a trivial extension of NLM to include the temporal dimension, bilateral filtering, anisotropic diffusion filtering, wavelet adaptive multiscale products threshold, and traditional wavelet thresholding-are presented. The evaluations include quantitative evaluations using simulated data and real data (20 DCE-MRI data sets from routine clinical breast MRI examinations) as well as qualitative evaluations using the same real data (24 observers: 14 image/signal-processing specialists, 10 clinical breast MRI radiographers). The results of the quantitative evaluation using the simulated data show that the DNLM algorithm consistently yields the smallest MSE between the denoised image and its corresponding original noiseless version. The results of the quantitative evaluation using the real data provide evidence, at the alpha = 0.05 level of significance, that the DNLM algorithm yields the smallest MSE between the denoised image and its corresponding original noiseless version. The results of the qualitative evaluation provide evidence, at the alpha = 0.05 level of significance, that the DNLM algorithm performs visually better than all of the other algorithms. Collectively the qualitative and quantitative results suggest that the DNLM algorithm more effectively attenuates noise in DCE MR images than any of the other algorithms.

Sketch based image coding decomposes an input image into a piece-wise smooth approximation image and a residual image. Image compression by gradient field integration follows this model, but differs by generating the approximation image from gradient data along edge contours and regularly sampled low resolution image data. This allows direct and efficient calculation of an approximation image which is smooth between edges. In this paper, we describe the image compression by gradient field integration approach, together with a low complexity implementation intended for near visually lossless compression of mixed content images. The implementation uses gradient field integration by scanline convolution, simple edge data extraction, DCT residual image coding and is block-based; it is suited to compression of images containing sharp edges occurring along pixel borders. Compression results are provided indicating potential gains from this method.

Virtually all existing classification techniques label one sample at a time. In this paper, we highlight the potential benefits of group based classification (GBC), where the classifier labels a group of homogeneous samples. In this way, GBC can take advantage of the additional prior knowledge that all samples belong to the same, unknown, class. We pose GBC in a generic hypothesis testing framework requiring the selection of an appropriate sample and test statistic. We then evaluate one simple example of GBC on both synthetic and real data sets and demonstrate that GBC may be a promising approach in applications where the test data can be arranged into homogenous subsets.

Fundamental matrix estimation is a central problem in computer vision and forms the basis of tasks such as stereo imaging and structure from motion. Existing algorithms typically analyze the relative geometries of matched feature points identified in both projected views. Automated feature matching is itself a challenging problem. Results typically have a large number of false matches. Traditional fundamental matrix estimation methods are very sensitive to matching errors, which led naturally to the application of robust statistical estimation techniques to the problem. In this work, an entirely novel approach is proposed to the fundamental matrix estimation problem. Instead of analyzing the geometry of matched feature points, the problem is recast in the frequency domain through the use of Integral Projection, showing how this is a reasonable model for orthographic cameras. The problem now reduces to one of identifying matching lines in the frequency domain which, most importantly, requires no feature matching or correspondence information. Experimental results on both real and synthetic data are presented that demonstrate the algorithm is a practical technique for fundamental matrix estimation. The behavior of the proposed algorithm is additionally characterized with respect to input noise, feature counts, and other parameters of interest.

In this paper, we propose a novel algorithm for rule extraction from support vector machines (SVMs), termed SQRex-SVM. The proposed method extracts rules directly from the support vectors (SVs) of a trained SVM using a modified sequential covering algorithm. Rules are generated based on an ordered search of the most discriminative features, as measured by interclass separation. Rule performance is then evaluated using measured rates of true and false positives and the area under the receiver operating characteristic (ROC) curve (AUC). Results are presented on a number of commonly used data sets that show the rules produced by SQRex-SVM exhibit both improved generalization performance and smaller more comprehensible rule sets compared to both other SVM rule extraction techniques and direct rule learning techniques.

One of the key problems in computer vision is the recovery of epipolar geometry constraints between different camera views. The majority of existing techniques rely on point correspondences, which are typically perturbed by mismatches and noise, hence limiting the accuracy of these techniques. To overcome these limitations, we propose a novel approach that estimates epipolar geometry constraints based on a statistical model in the RADON domain. The method requires no correspondences, explicit constraints on the data or assumptions regarding the scene structure. Results are presented on both synthetic and real data that show the method's robustness to noise and outliers.

In this paper, we describe a model of the human visual system (HVS) based on the wavelet transform. This model is largely based on a previously proposed model, but has a number of modifications that make it more amenable to potential integration into a wavelet based image compression scheme. These modifications include the use of a separable wavelet transform instead of the cortex transform, the application of a wavelet contrast sensitivity function (CSF), and a simplified definition of subband contrast that allows one to predict the noise visibility directly from the wavelet coefficients. Initially, we outline the luminance, frequency, and masking sensitivities of the HVS and discuss how these can be incorporated into the wavelet transform. We then outline a number of limitations of the wavelet transform as a model of the HVS, namely the lack of translational invariance and poor orientation sensitivity. In order to investigate the efficacy of this wavelet based model, a wavelet visible difference predictor (WVDP) is described. The WVDP is then used to predict visible differences between an original and compressed (or noisy) image. Results are presented to emphasize the limitations of commonly used measures of image quality and to demonstrate the performance of the WVDP. The paper concludes with suggestions on how the WVDP can be used to determine a visually optimal quantization strategy for wavelet coefficients and produce a quantitative measure of image quality.

Proposes a rule-based inductive learning algorithm called multiscale classification (MSC). It can be applied to any N-dimensional real or binary classification problem to classify the training data by successively splitting the feature space in half. The algorithm has several significant differences from existing rule-based approaches: learning is incremental, the tree is non-binary, and backtracking of decisions is possible to some extent. The paper first provides background on current machine learning techniques and outlines some of their strengths and weaknesses. It then describes the MSC algorithm and compares it to other inductive learning algorithms with particular reference to ID3, C4.5, and back-propagation neural networks. Its performance on a number of standard benchmark problems is then discussed and related to standard learning issues such as generalization, representational power, and over-specialization.