Telmo Amaral

Using Different Cost Functions to Train Stacked Auto-Encoders

Deep neural networks comprise several hidden layers of units, which can be pre-trained one at a time via an unsupervised greedy approach. A whole network can then be trained (fine-tuned) in a supervised fashion. One possible pre-training strategy is to regard each hidden layer in the network as the input layer of an auto-encoder. Since auto-encoders aim to reconstruct their own input, their training must be based on some cost function capable of measuring reconstruction performance. Similarly, the supervised fine-tuning of a deep network needs to be based on some cost function that reflects prediction performance. In this work we compare different combinations of cost functions in terms of their impact on layer-wise reconstruction performance and on supervised classification performance of deep networks. We employed two classic functions, namely the cross-entropy (CE) cost and the sum of squared errors (SSE), as well as the exponential (EXP) cost, inspired by the error entropy concept. Our results were based on a number of artificial and real-world data sets.

Immunohistochemical analysis of breast tissue microarray images using contextual classifiers

Background: Tissue microarrays (TMAs) are an important tool in translational research for examining multiple cancers for molecular and protein markers. Automatic immunohistochemical (IHC) scoring of breast TMA images remains a challenging problem. Methods: A two-stage approach that involves localization of regions of invasive and in-situ carcinoma followed by ordinal IHC scoring of nuclei in these regions is proposed. The localization stage classifies locations on a grid as tumor or non-tumor based on local image features. These classifications are then refined using an auto-context algorithm called spin-context. Spin-context uses a series of classifiers to integrate image feature information with spatial context information in the form of estimated class probabilities. This is achieved in a rotationally-invariant manner. The second stage estimates ordinal IHC scores in terms of the strength of staining and the proportion of nuclei stained. These estimates take the form of posterior probabilities, enabling images with uncertain scores to be referred for pathologist review. Results: The method was validated against manual pathologist scoring on two nuclear markers, progesterone receptor (PR) and estrogen receptor (ER). Errors for PR data were consistently lower than those achieved with ER data. Scoring was in terms of estimated proportion of cells that were positively stained (scored on an ordinal scale of 0-6) and perceived strength of staining (scored on an ordinal scale of 0-3). Average absolute differences between predicted scores and pathologist-assigned scores were 0.74 for proportion of cells and 0.35 for strength of staining (PR). Conclusions: The use of context information via spin-context improved the precision and recall of tumor localization. The combination of the spin-context localization method with the automated scoring method resulted in reduced IHC scoring errors.

Classification and Immunohistochemical Scoring of Breast Tissue Microarray Spots

Tissue microarrays (TMAs) facilitate the survey of very large numbers of tumors. However, the manual assessment of stained TMA sections constitutes a bottleneck in the pathologists work flow. This paper presents a computational pipeline for automatically classifying and scoring breast cancer TMA spots that have been subjected to nuclear immunostaining. Spots are classified based on a bag of visual words approach. Immunohistochemical scoring is performed by computing spot features reflecting the proportion of epithelial nuclei that are stained and the strength of that staining. These are then mapped onto an ordinal scale used by pathologists. Multilayer perceptron classifiers are compared with latent topic models and support vector machines for spot classification, and with Gaussian process ordinal regression and linear models for scoring. Intraobserver variation is also reported. The use of posterior entropy to identify uncertain cases is demonstrated. Evaluation is performed using TMA images stained for progesterone receptor.

Classification of breast-tissue microarray spots using colour and local invariants

Breast tissue microarrays facilitate the survey of very large numbers of tumours but their scoring by pathologists is time consuming, typically highly quantised and not without error. Automated segmentation of cells and intra-cellular compartments in such data can be problematic for reasons that include cell overlapping, complex tissue structure, debris, and variable appearance. This paper proposes a computationally efficient approach that uses colour and differential invariants to assign class posterior probabilities to pixels and then performs probabilistic classification of TMA spots using features analogous to the Quickscore system currently used by pathologists. It does not rely on accurate segmentation of individual cells. Classification performance at both pixel and spot levels was assessed using 110 spots from the adjuvant breast cancer (ABC) chemotherapy trial. The use of differential invariants in addition to colour yielded a small improvement in accuracy. Some reasons for classification results in disagreement with pathologist-provided labels are discussed and include noise in the class labels.

Lifestyle monitoring as a predictive tool in telecare

Six people with multiple health problems living in an extra care housing scheme for older people with vision impairment agreed to take part in a telecare trial. An average of 14.8 sensors was installed in each of the flats. The monitoring period began in January 2006 and lasted for 10 months. The data acquired by the sensors installed in each flat was transmitted from the home unit to a central computer. Four interviews were conducted with each participant. Halfway through the study we created a number of case studies of sensor activity at the time of known events in the lives of the participants, together with an attempt to interpret these patterns of activity with the benefit of available contextual information. Although our investigation showed that sensors are capable of identifying some changes in daily routines at the time of important events, the interpretation of such changes requires a large amount of contextual information and the involvement of participants themselves. Various technical and operational difficulties will need to be resolved before it will be possible to use lifestyle monitoring predictively.

Speeching: Mobile Crowdsourced Speech Assessment to Support Self-Monitoring and Management for People with Parkinson's

We present Speeching, a mobile application that uses crowdsourcing to support the self-monitoring and management of speech and voice issues for people with Parkinsons (PwP). The application allows participants to audio record short voice tasks, which are then rated and assessed by crowd workers. Speeching then feeds these results back to provide users with examples of how they were perceived by listeners unconnected to them (thus not used to their speech patterns). We conducted our study in two phases. First we assessed the feasibility of utilising the crowd to provide ratings of speech and voice that are comparable to those of experts. We then conducted a trial to evaluate how the provision of feedback, using Speeching, was valued by PwP. Our study highlights how applications like Speeching open up new opportunities for self-monitoring in digital health and wellbeing, and provide a means for those without regular access to clinical assessment services to practice and get meaningful feedback on their speech.

Segmentation of organs in pig offal using auto-context

The segmentation of 2D images of 3D non-rigid objects into their constituent parts can pose challenging problems, such as missing and occluded parts, weak constraints over the spatial arrangement of parts, and variance in form and appearance. These problems have been addressed via segmentation methods that incorporate spatial context information, such as the auto-context technique. In this paper, we address for the first time the problem of segmenting multiple organs in images of pig offal, a challenging image analysis task that constitutes an essential step towards automated screening at abattoir for signs of sub-clinical diseases. We applied auto-context segmentation to a large data set of images and explored the effect of complementing conventional context features with integral features suited to our application.

Improving Performance on Problems with Few Labelled Data by Reusing Stacked Auto-Encoders

Deep architectures have been used in transfer learning applications, with the aim of improving the performance of networks designed for a given problem by reusing knowledge from another problem. In this work we addressed the transfer of knowledge between deep networks used as classifiers of digit and shape images, considering cases where only the set of class labels, or only the data distribution, changed from source to target problem. Our main goal was to study how the performance of knowledge transfer between such problems would be affected by varying the number of layers being retrained and the amount of data used in that retraining. Generally, reusing networks trained for a different label set led to better results than reusing networks trained for a different data distribution. In particular, reusing for less classes a network trained for more classes was beneficial for virtually any amount of training data. In all cases, retraining only one layer to save time consistently led to poorer performance. The results obtained when retraining for upright digits a network trained for rotated digits raise the hypothesis that transfer learning could be used to better deal with image classification problems in which only a small amount of labelled data is available for training.

Transfer Learning Using Rotated Image Data to Improve Deep Neural Network Performance

In this work we explore the idea that, in the presence of a small training set of images, it could be beneficial to use that set itself to obtain a transformed training set (by performing a random rotation on each sample), train a source network using the transformed data, then retrain the source network using the original data. Applying this transfer learning technique to three different types of character data, we achieve average relative improvements between 6 % and 16 % in the classification test error. Furthermore, we show that it is possible to achieve relative improvements between 8 % and 42 % in cases where the amount of original training samples is very limited (30 samples per class), by introducing not just one rotation but several random rotations per sample.

CrowdEyes: crowdsourcing for robust real-world mobile eye tracking

Current eye tracking technologies have a number of drawbacks when it comes to practical use in real-world settings. Common challenges, such as high levels of daylight, eyewear (e.g. spectacles or contact lenses) and eye make-up, give rise to noise that undermines their utility as a standard component for mobile computing, design, and evaluation. To work around these challenges, we introduce CrowdEyes, a mobile eye tracking solution that utilizes crowdsourcing for increased tracking accuracy and robustness. We present a pupil detection task design for crowd workers together with a study that demonstrates the high-level accuracy of crowdsourced pupil detection in comparison to state-of-the-art pupil detection algorithms. We further demonstrate the utility of our crowdsourced analysis pipeline in a fixation tagging task. In this paper, we validate the accuracy and robustness of harnessing the crowd as both an alternative and complement to automated pupil detection algorithms, and explore the associated costs and quality of our crowdsourcing approach.