BIOIMAGING 2020 Abstracts

Full Papers Short Papers
Full Papers
Paper Nr: 1
Title:

Photoluminescent Imaging for an Effective Cancer Diagnosis using Upconversion Nanoparticles
Authors:

Rafia Rafique and Tae J. Park
Abstract: Advancements in the synthesis of upconversion nanoparticles (UCNP) can enable a broad range of biomedical applications. Herein, we fabricated NaYF4:Yb3+/Er3+ UCNP and polyacrylic acid conjugated UCNPs (UCNP@PAA). Characterizations of the resulting particles were conducted using electron microscopy and spectroscopy, X-ray diffraction (XRD), and upconversion luminescence (UCL) analysis. We demonstrated that particles were synthesized with good homogeneity, hexagonal phase, and UCL efficiency. The UCNP@PAA maintained their original particle size and luminescence properties, cellular nontoxicity, in vitro bioimaging, and biocompatibility. Based on these results, we suggest that these particles can be applied in drug-delivery systems and as bioimaging agents in the future.
Download
Paper Nr: 7
Title:

Assessment of Gallbladder Wall Vascularity from Laparoscopic Images using Deep Learning
Authors:

Constantinos Loukas and Dimitrios Schizas
Abstract: Despite the significant progress in content-based video analysis of surgical procedures, methods on analyzing still images acquired during the operation are limited. In this paper we elaborate on a novel idea for computer vision-based assessment of the vascularity of the gallbladder (GB) wall, using frames extracted from videos of laparoscopic cholecystectomy. The motivation was based on the fact that the wall’s vascular pattern provides an indirect indication of the GB condition (e.g. fat coverage, wall thickening, inflammation), which in turn is usually related to the operation complexity. As the GB wall vascularity may appear irregular, in this study we focus on the classification of rectangular sub-regions (patches). A convolutional neural network (CNN) is proposed for patch classification based on two ground-truth annotation schemes: 3-classes (Low, Medium and High vascularity) and 2-classes (Low vs. High). Moreover, we employed three popular classifiers with a rich set of hand-crafted descriptors. The CNN achieved the best performance with accuracy: 98% and 83.1%, and mean F1-score: 98% and 80.4%, for 2-class and 3-class classification, respectively. The other methods’ performance was lower by 2%-6% (2-classes) and 6%-17% (3 classes). Our results indicate that CNN-based patch classification is promising for intraoperative assessment of the GB wall vascularity.
Download
Paper Nr: 13
Title:

Curriculum Deep Reinforcement Learning with Different Exploration Strategies: A Feasibility Study on Cardiac Landmark Detection
Authors:

Patricio Astudillo, Peter Mortier, Matthieu De Beule and Francis Wyffels
Abstract: Transcatheter aortic valve implantation (TAVI) is associated with conduction abnormalities and the mechanical interaction between the prosthesis and the atrioventricular (AV) conduction path cause these life-threatening arrhythmias. Pre-operative assessment of the location of the AV conduction path can help to understand the risk of post-TAVI conduction abnormalities. As the AV conduction path is not visible on cardiac CT, the inferior border of the membranous septum can be used as an anatomical landmark. Detecting this border automatically, accurately and efficiently would save operator time and thus benefit pre-operative planning. This preliminary study was performed to identify the feasibility of 3D landmark detection in cardiac CT images with curriculum deep Q-learning. In this study, curriculum learning was used to gradually teach an artificial agent to detect this anatomical landmark from cardiac CT. This agent was equipped with a small field of view and burdened with a large action-space. Moreover, we introduced two novel action-selection strategies: α-decay and action-dropout. We compared these two strategies to the already established ε-decay strategy and observed that α-decay yielded the most accurate results. Limited computational resources were used to ensure reproducibility. In order to maximize the amount of patient data, the method was cross-validated with k-folding for all three action-selection strategies. An inter-operator variability study was conducted to assess the accuracy of the method.
Download
Paper Nr: 14
Title:

Automated 3D Labelling of Fibroblasts and Endothelial Cells in SEM-Imaged Placenta using Deep Learning
Authors:

Benita S. Mackay, Sophie Blundell, Olivia Etter, Yunhui Xie, Michael T. McDonnel, Matthew Praeger, James Grant-Jacob, Robert Eason, Rohan Lewis and Ben Mills
Abstract: Analysis of fibroblasts within placenta is necessary for research into placental growth-factors, which are linked to lifelong health and chronic disease risk. 2D analysis of fibroblasts can be challenging due to the variation and complexity of their structure. 3D imaging can provide important visualisation, but the images produced are extremely labour intensive to construct because of the extensive manual processing required. Machine learning can be used to automate the labelling process for faster 3D analysis. Here, a deep neural network is trained to label a fibroblast from serial block face scanning electron microscopy (SBFSEM) placental imaging.
Download
Paper Nr: 20
Title:

Low-density EEG for Source Activity Reconstruction using Partial Brain Models
Authors:

Andres F. Soler, Eduardo Giraldo and Marta Molinas
Abstract: Brain mapping studies have shown that the source reconstruction performs with high accuracy by using high-density EEG montages, however, several EEG devices in the market provide low-density configurations and thus source reconstruction is considered out of the scope of those devices. In this work, our aim is to use a few numbers of electrodes to reconstruct the neural activity using partial brain models, therefore, we presented a pipeline to estimate the brain activity using a low-density EEG on brain regions of interest, the partial brain model formulation and several criteria for channel selection. Two regions have been considered to be studied, the occipital region and motor cortex region. For the presented study synthetic EEG signals were generated simulating the activation of sources with a frequency in the beta range at the occipital region, and mu rhythm range at the motor cortex areas. Novel methods for electrode reduction and models for specific brain areas are presented. We assessed the quality of the reconstructions by measuring the localization error, obtaining a mean localization error below 7 mm and 16 mm with sLORETA and MSP methods respectively, by using a low-density EEG with eight channels and partial brain models.
Download
Paper Nr: 21
Title:

Fully Automated Volumetric Measurement of Malignant Pleural Mesothelioma from Computed Tomography Images by Deep Learning: Preliminary Results of an Internal Validation
Authors:

Owen Anderson, Andrew C. Kidd, Keith A. Goatman, Alexander J. Weir, Jeremy Voisey, Vismantas Dilys, Jan P. Siebert and Kevin G. Blyth
Abstract: Malignant Pleural Mesothelioma (MPM) is a cancer associated with prior exposure to asbestos fibres. Unlike most tumours, which are roughly spherical, MPM grows like a rind surrounding the lung. This irregular shape poses significant clinical and technical challenges. Accurate tumour measurements are necessary to determine treatment efficacy, but manual segmentation is tedious, time-consuming and associated with high intra- and inter-observer variation. In addition, uncertainty is compounded by poor differentiation in the computed tomography (CT) image between MPM and other common features. We describe herein an internal validation of a fully automatic tool to generate volumetric segmentations of MPM tumours using a convolutional neural network (CNN). The system was trained using the first 123 CT volumetric datasets from a planned total of 403 scans. Each scan was manually segmented to provide the expert ground truth. Evaluation was by seven-fold cross validation on a subset of 80/123 datasets that have full volumetric segmentations. The mean volume of MPM tumour in these datasets is 405.1 cm3 (standard deviation 271.5 cm3). Following three-dimensional binary closing of the manual annotations to improve inter-slice consistency, the mean volume difference between the manual and automatic measurements is 27.2 cm3, which is not significantly different from zero difference (p = 0:225). The 95% limits of agreement between the manual and automated measurements are between -417 and +363 cm3. The mean Dice overlap coefficient was 0.64, which is comparable with inter-observer measurements reported elsewhere. To our knowledge, this is the first algorithm of its kind that fully automates and evaluates measurement of the MPM tumour volume. The next step will be to evaluate the method on the remaining unseen multi-centre evaluation set. Such an algorithm has possible future application to pharmaceutical trials (where it offers a repeatable study end point) and to routine care (where it allows tumour progression to be assessed rapidly to enhance therapeutic clinical decision making).
Download
Paper Nr: 22
Title:

BADRESC: Brain Anomaly Detection based on Registration Errors and Supervoxel Classification
Authors:

Samuel B. Martins, Alexandre X. Falcão and Alexandru C. Telea
Abstract: Automatic detection of brain anomalies in MR images is very challenging and complex due to intensity similarity between lesions and normal tissues as well as the large variability in shape, size, and location among different anomalies. Inspired by groupwise shape analysis, we adapt a recent fully unsupervised supervoxel-based approach (SAAD) — designed for abnormal asymmetry detection of the hemispheres — to detect brain anomalies from registration errors. Our method, called BADRESC, extracts supervoxels inside the right and left hemispheres, cerebellum, and brainstem, models registration errors for each supervoxel, and treats outliers as anomalies. Experimental results on MR-T1 brain images of stroke patients show that BADRESC attains similar detection rate for hemispheric lesions in comparison to SAAD with substantially less false positives. It also presents promising detection scores for lesions in the cerebellum and brainstem.
Download
Paper Nr: 35
Title:

Glioma Grade Classification via Omics Imaging
Authors:

Lucia Maddalena, Ilaria Granata, Ichcha Manipur, Mario Manzo and Mario R. Guarracino
Abstract: Omics imaging is an emerging interdisciplinary field concerned with the integration of data collected from biomedical images and omics experiments. Bringing together information coming from different sources, it permits to reveal hidden genotype-phenotype relationships, with the aim of better understanding the onset and progression of many diseases, and identifying new diagnostic and prognostic biomarkers. In this work, we present an omics imaging approach to the classification of different grades of gliomas, which are primary brain tumors arising from glial cells, as this is of critical clinical importance for making decisions regarding initial and subsequent treatment strategies. Imaging data come from analyses available in The Cancer Imaging Archive, while omics attributes are extracted by integrating metabolic models with transcriptomic data available from the Genomic Data Commons portal. We investigate the results of feature selection for the two types of data separately, as well as for the integrated data, providing hints on the most distinctive ones that can be exploited as biomarkers for glioma grading. Moreover, we show how the integrated data can provide additional clinical information as compared to the two types of data separately, leading to higher performance. We believe our results can be valuable to clinical tests in practice.
Download
Short Papers
Paper Nr: 5
Title:

Comparison of Gadolinium Contrast Agent Retention in Patients Receiving Multiple Contrast-enhanced MRI Exams
Authors:

Ryan Fisher, Vikas Jain, Jonathan Glaab and Aubrey McMillan
Abstract: Gadolinium-based contrast agents have long been utilized in magnetic resonance imaging (MRI) to enhance image quality. Aside from the few reported cases of Nephrogenic Systemic Fibrosis in patients with severely compromised renal function, these contrast agents have generally been viewed as safe. However, recent studies have shown evidence of the retention of potentially toxic gadolinium well beyond the previously recognized clearing times in patients with normal renal function. This retention has been shown via persistent hyper-intense signal in certain brain regions in unenhanced MRI exams. The exact form of retained gadolinium and its long-term potential health effects remain unknown at this time. Due to concerns over retained gadolinium, our hospital switched to a more stably bound contrast agent in the spring of 2018. This study examined brain MRI images from patients with multiple contrast-enhanced exams using either the older, more unstable, linear agent, and the newer, more stable, macrocyclic agent. Signal intensities were measured in the globus pallidus and dentate nucleus; regions of the brain that have previously been shown to accumulate heavy metals such as gadolinium. Statistically significant increases in signal intensity were seen in the dentate nucleus in the linear contrast agent group, but not in the macrocyclic agent group. No significant signal increases were seen with either agent in the globus pallidus region of the brain. No correlation was seen between signal increase and the volume of contrast agent administered for either region or contrast agent.
Download
Paper Nr: 6
Title:

Exploiting Bilateral Symmetry in Brain Lesion Segmentation with Reflective Registration
Authors:

Kevin Raina, Uladzimir Yahorau and Tanya Schmah
Abstract: Brain lesions, including stroke lesions and tumours, have a high degree of variability in terms of location, size, intensity and form, making automatic segmentation difficult. We propose an improvement to existing segmentation methods by exploiting the bilateral quasi-symmetry of healthy brains, which breaks down when lesions are present. Specifically, we use nonlinear registration of a neuroimage to a reflected version of itself (“reflective registration”) to determine for each voxel its homologous (corresponding) voxel in the other hemisphere. A patch around the homologous voxel is added as a set of new features to the segmentation algorithm. To evaluate this method, we implemented two different CNN-based multimodal MRI stroke lesion segmentation algorithms, and then augmented them by adding extra symmetry features using the reflective registration method described above. For each architecture, we compared the performance with and without symmetry augmentation, on the SISS Training dataset of the Ischemic Stroke Lesion Segmentation Challenge (ISLES) 2015 challenge. Using linear reflective registration improves performance over baseline, but nonlinear reflective registration gives significantly better results: an improvement in Dice coefficient of 13 percentage points over baseline for one architecture and 9 points for the other. We argue for the broad applicability of adding symmetric features to existing segmentation algorithms, specifically using the proposed nonlinear, template-free method.
Download
Paper Nr: 8
Title:

Extraction of Intrinsic Fluorescence in Fluorescence Imaging of Turbid Tissues
Authors:

Gennadi Saiko and Alexandre Douplik
Abstract: Interpretation of tissue fluorescence spectra can be complicated due to interplay with tissue optics. We have developed a photon propagation approach for correction of fluorescence on absorption in two realistic scenarios: when fluorophores are located a) on the surface of the turbid tissue and b) in a layer inside the turbid tissue. The approach takes into account the diffuse reflection of the tissue at excitation and emission wavelengths and does not require any precise measurement of optical properties (e.g., coefficient of absorption). The approach can be implemented using an inexpensive imaging setup and can be used in any setting.
Download
Paper Nr: 9
Title:

Water-sensitive Gelatin Phantoms for Skin Water Content Imaging
Authors:

Gennadi Saiko and Alexandre Douplik
Abstract: Oxygen supply to tissues can be seriously impacted during wound healing. Edema (accumulation of fluids in interstitial space) can increase the distance between capillaries, thus decreasing oxygen supply to cells. There is no standard clinical tool for quantification of edema, and early edema detection (preferably preclinical) is of great clinical need. Multispectral imaging can be a helpful clinical tool to characterize water content in the skin. However, to develop and validate this technology, a reliable water-sensitive preclinical model has to be developed. The scope of this work is to develop a water-responsive skin model and assess the feasibility of extracting water content using multispectral imaging. Methods: A phantom fabrication protocol has been developed. The phantoms are based on the gelatin crosslinked with glutaraldehyde. TiO2 nanoparticles were added to mimic the optical properties of the skin. To emulate various water content, phantoms were dipped in the water for various duration. The phantoms were imaged using the Multi-Spectral Imaging Device (MSID) (Swift Medical Inc, Toronto). MSID is a multispectral imaging system for visualization of tissue chromophores in surface tissues. It uses 12-bit scientific-grade NIR-enhanced monochrome camera (Basler, Germany) and ten wavelength light source (600-1000nm range) to visualize the distribution of oxy-, deoxyhemoglobins, methemoglobin, water, and melanin. The imaging distance is 30cm, the field of view: 7x7cm. Results: Initial results show that the developed model mimics the optical scattering properties of the skin. MSID was able to extract water content using a full set (ten wavelengths) and a subset (three wavelengths) of channels. Conclusions: A new water responsive model for skin moisture imaging has been developed. Initial experiments with multispectral imaging of these phantoms show feasibility of tissue water content imaging with Si-based cameras.
Download
Paper Nr: 12
Title:

PySpot: A Python based Framework for the Assessment of Laser-modified 3D Microstructures for Windows and Raspbian
Authors:

Hannah Janout, Bianca Buchegger, Andreas Haghofer, Dominic Hoeglinger, Jaroslaw Jacak, Stephan Winkler and Armin Hochreiner
Abstract: Biocompatible 3D microstructures created with laser lithography and modified for optimal cell growth with laser grafting, can imitate the 3D structure cells naturally grow in. For the evaluation of the quality and success of those 3D microstructures, specialized software is required. Our software PySpot can load 2D and 3D images and analyze those through image segmentation, edge detection, and surface plots. Additionally, the creation and modification of regions of interest (ROI) allow for the quality evaluation of specific areas in an image by intensity analysis. 3D rendering allows for identifying complex geometrical properties. Furthermore, PySpot is executable on Windows as well as on Raspbian, which makes it flexible to use.
Download
Paper Nr: 15
Title:

HEp-2 Intensity Classification based on Deep Fine-tuning
Authors:

Vincenzo Taormina, Donato Cascio, Leonardo Abbene and Giuseppe Raso
Abstract: The classification of HEp-2 images, conducted through Indirect ImmunoFluorescence (IIF) gold standard method, in the positive / negative classes, is the first step in the diagnosis of autoimmune diseases. Since the test is often difficult to interpret, the research world has been looking for technological features for this problem. In recent years the methods of deep learning have overcome the other machine learning techniques in their effectiveness and robustness, and now they prevail in artificial intelligence studies. In this context, CNNs have played a significant role especially in the biomedical field. In this work we analysed the capabilities of CNN for fluorescence classification of HEp-2 images. To this end, the GoogLeNet pre-trained network was used. The method was developed and tested using the public database A.I.D.A. For the analysis of pre-trained network, the two strategies were used: as features extractors (coupled with SVM classifiers) and after fine-tuning. Performance analysis was conducted in terms of ROC (Receiver Operating Characteristic) curve. The best result obtained with the fine-tuning method showed an excellent ability to discriminate between classes, with an area under the ROC curve (AUC) of 98.4% and an accuracy of 93%. The classification result using the CNN as features extractor obtained 97.3% of AUC, showing a difference in performance between the two strategies of little significance.
Download
Paper Nr: 19
Title:

On Benchmarking Cell Nuclei Segmentation Algorithms for Fluorescence Microscopy
Authors:

Frederike Wirth, Eva-Maria Brinkmann and Klaus Brinker
Abstract: Cell nuclei detection is an essential step in the context of many image analysis tasks related to microscopy images and therefore also plays a role in highly topical fields of research like the development of personalised immunotherapy approaches against several types of cancer. Motivated by this observation, a whole zoo of advanced methods to accomplish cell nuclei segmentation has been proposed in recent years. This development in turn stresses the need to set up a well-justified and reproducible standard of comparison for the evaluation of these sophisticated approaches. In this paper, we describe how such a reference framework based on standard seeded watershed segmentation for fully automatic cell nuclei detection can be obtained. In particular, we provide a detailed review of a publicly available dataset, give a detailed account of the methods and evaluation measures we consider to enable the highest possible reproducibility of our results and discuss the suitability of different variants of seeded watershed segmentation for the mentioned purposes.
Download
Paper Nr: 27
Title:

Modelling Brain Lesion Volume in Patches with CNN-based Poisson Regression
Authors:

Kevin Raina
Abstract: Monitoring the progression of lesions is important for clinical response. Summary statistics such as lesion volume are objective and easy to interpret, which can help clinicians assess lesion growth or decay. CNNs are commonly used in medical image segmentation for their ability to produce useful features within large contexts and their associated efficient iterative patch-based training. Many CNN architectures require hundreds of thousands parameters to yield a good segmentation. In this work, an efficient, computationally inexpensive CNN is implemented to estimate the number of lesion voxels in a predefined patch size from magnetic resonance (MR) images. The output of the CNN is interpreted as the conditional Poisson parameter over the patch, allowing standard mini-batch gradient descent to be employed. The ISLES2015 (SISS) data is used to train and evaluate the model, which by estimating lesion volume from raw features, accurately identified the lesion image with the larger lesion volume for 86% of paired sample patches. An argument for the development and use of estimating lesion volumes to also aid in model selection for segmentation is made.
Download
Paper Nr: 29
Title:

Detection and Categorisation of Multilevel High-sensitivity Cardiovascular Biomarkers from Lateral Flow Immunoassay Images via Recurrent Neural Networks
Authors:

Min Jing, Donal McLaughlin, David Steele, Sara McNamee, Brian MacNamee, Patrick Cullen, Dewar Finlay and James McLaughlin
Abstract: Lateral Flow Immunoassays (LFA) have the potential to provide low cost, rapid and highly efficacious Point-of-Care (PoC) diagnostic testing in resource limited settings. Traditional LFA testing is semi-quantitative based on the calibration curve, which faces challenges in the detection of multilevel high-sensitivity biomarkers due its low sensitivity. This paper proposes a novel framework in which the LFA images are acquired from a designed CMOS reader system under controlled lighting. Unlike most existing approaches based on image intensity, the proposed system does not require detection of region of interest (ROI), instead each row of the LFA image was considered as time series signals. The Long Short-Term Memory (LSTM) network was deployed to classify the LFA data obtained from cardiovascular biomarker, C-Reactive Protein (CRP), at eight concentration levels (within the range 0-5mg/L) that are aligned with clinically actionable categories. The performance under different arrangements for input dimension and parameters were evaluated. The preliminary results show that the proposed LSTM outperforms other popular classification methods, which demonstrate the capability of the proposed system to detect high-sensitivity CRP and suggests the potential of applications for early risk assessment of cardiovascular diseases (CVD).
Download
Paper Nr: 31
Title:

Image Quality Comparison between Synthetic 2D Mammograms Obtained with 15º and 40º X-ray Tube Angular Range: A Quantitative Phantom Study
Authors:

R. Lamastra, P. Barca, M. G. Bisogni, D. Caramella, V. Rosso, R. M. Tucciariello, A. C. Traino and M. E. Fantacci
Abstract: In this work we present an image quality comparison between synthesized mammograms (SMs) obtained from Digital Breast Tomosynthesis (DBT) acquisitions with 15° (SM15) and 40° (SM40) X-ray tube angular range. In fact, since wide-angle DBT is characterized by a better spatial resolution in depth but also by worse performance in detecting microcalcifications than narrow-angle DBT, an objective image quality analysis of SM images could be of pratical interest. Four phantoms were employed in this study and their images were acquired using an Amulet Innovality mammographic device. The image quality comparison was conducted by evaluating spatial resolution, contrast and noise properties of the images. Our results show that SM40 images are characterized by better spatial resolution performance than SM15 in terms of Modulation Transfer Function but also by worse performance in the detection of low-contrast details. In fact, higher contrast-to-noise ratio values were obtained with SM15 than with SM40. Noise properties of the images were also investigated through the Noise Power Spectrum (NPS) calculation: no differences in NPS shapes were found in both modalities, while noise magnitude results significantly different. In addition, Signal-to-Noise Ratio (SNR) spatial distribution evaluation was assessed by computing SNR maps, in which different pattern were observed.
Download
Paper Nr: 33
Title:

The Behaviour of Neuro-2A Cells on Silicon Substrates with Various Topographies Generated by Femtosecond Laser Micromachining
Authors:

Sara Mingu, Ihor Pavlov, Çağdaş D. Son and Alpan Bek
Abstract: The interaction of neural cells with silicon surfaces is important for basic research as well as for various possible applications, such as silicon-based neural implants and neurochips. Laser structuring of silicon provides a quick and versatile method for the generation of complex, hierarchical topographies on precise locations of the substrate. The behaviour of Neuro-2A cells with laser-structured silicon substrates was studied using a live-imaging setup with fluorescence microscopy. Neuro-2A cells were able to adhere to polished silicon, ripples and microcolumns to different extents, depending on the substrate topography and incubation time. Initially, cells adhere much better to structured areas, resulting in visible cell patterning on the substrates. Time-lapse microscopy revealed cell exploration and motility behaviours on the substrates. Cell motility was significantly decreased on structured substrates, with whole area microcolumns having the slowest cell motility. On polished silicon, cells were found to interact with the substrates using lamellipodia and filopodia. After 24 or 48 hours, cells were better able to adhere to polished as well as structured silicon. Neurite alignment was observed on microcolumn and trench substrates. On the other hand, highly processed substrates were inhibitory to cell growth and resulted in poor cell health.
Download
Paper Nr: 36
Title:

Artificial Neural Networks for Quantitative Microwave Breast Imaging
Authors:

M. Ambrosanio, S. Franceschini, F. Baselice and V. Pascazio
Abstract: This paper is focused on the use of artificial neural networks (ANNs) for biomedical microwave imaging of breast tissues in the framework of advanced breast cancer imaging techniques. The proposed scheme processes the scattered field collected at receivers locations of a multiview-multistatic system and aims at providing an estimate of the morphological and dielectric features of the breast tissues, which represents a strongly nonlinear scenario with several challenging aspects. In order to train the network, a simulated data set has been created by implementing the forward problem and an automatic randomly-shaped breast profile generator based on the statistical distribution of complex permittivity of breast biological tissues was developed. Some numerical tests were carried out to evaluate the performance of the proposed method and, in conclusion, we found that the use of ANNs for quantitative biomedical imaging purposes seems to be very promising.
Download
Paper Nr: 42
Title:

Evaluating Deep Learning Uncertainty Measures in Cephalometric Landmark Localization
Authors:

Dušan Drevický and Oldřich Kodym
Abstract: Cephalometric analysis is a key step in the process of dental treatment diagnosis, planning and surgery. Localization of a set of landmark points is an important but time-consuming and subjective part of this task. Deep learning is able to automate this process but the model predictions are usually given without any uncertainty information which is necessary in medical applications. This work evaluates three uncertainty measures applicable to deep learning models on the task of cephalometric landmark localization. We compare uncertainty estimation based on final network activation with an ensemble-based and a Bayesian-based approach. We conduct two experiments with elastically distorted cephalogram images and images containing undesirable horizontal skull rotation which the models should be able to detect as unfamiliar and unsuitable for automatic evaluation. We show that all three uncertainty measures have this detection capability and are a viable option when landmark localization with uncertainty estimation is required.
Download
Paper Nr: 3
Title:

Segmentation of Diabetic Retinopathy Lesions by Deep Learning: Achievements and Limitations
Authors:

Pedro Furtado
Abstract: Analysis of Eye Fundus Images (EFI) allows early diagnosis and grading of Diabetic Retinopathy (DR), detecting micro-aneurisms, exudates, haemorrhages, neo-vascularizations and other signs. Automated detection of individual lesions helps visualizing, characterizing and determining degree of DR. Today modified deep convolution neural networks (DCNNs) are state-of-the-art in most segmentation tasks. But the task of segmenting lesions in EFI is challenging due to sizes, varying shapes, similarity and lack of contrast with other parts of the EFI, so that the results are ambiguous. In this paper we test two DCNNs to do a preliminary evaluation of the strengths and limitations using publicly available data. We already conclude that the accuracies are good but the segmentations still have relevant deficiencies. Based on this, we identify the need for further assessment and suggest future work to improve segmentation approaches.
Download
Paper Nr: 4
Title:

Food Recognition: Can Deep Learning or Bag-of-Words Match Humans?
Authors:

Pedro Furtado
Abstract: Automated smartphone-based food recognition is a useful basis for applications targeted at dietary assessment. Dish recognition is a necessary step in that process. One of the possible approaches to use is deep learning-based recognition, another one is bag-of-words based classification. Deep learning has increasingly become the preferred approach to use in either this or other image classification tasks. Additionally, if humans are better recognizing the dish, the automated approach is useless (it will be less error-prone for the user to identify the dish instead of capturing the photo). We compare the alternatives of Deep Learning (DL), Bag-of-words (BoW) and Humans (H). The best deep learner beats humans when on few food categories, but looses if it has to learn many more food categories, which is expected in real contexts. We describe the approaches, analyze the results, draw conclusions and design further work to evaluate further and improve the approaches.
Download
Paper Nr: 16
Title:

3D Printed Human Foot Splint, Designed from MRI of the Luffa Cylindrica Dried Fruit
Authors:

Sergio Cerón-Escutia, Axayácatl Morales-Guadarrama, Silvia B. González-Brambila and David Vidal-García
Abstract: This article presents a method to design a splint from the MRI (Magnetic Resonance Imaging) of the dried fruit of a tropical plant called Luffa (Luffa cylindrica (L.) M.Roem), which shows the possibility of using synthetic forms in Nature to reproduce and apply them in the development of products, a concept known as bio-design. This fruit –similar to a cucumber-, when dried it becomes a fibrous and imbricated structure that confers interesting stiffness and lightness properties, which were used to design a splint for a human foot, different from the conventional plaster, through a reverse engineering process. Such structure was copied, with the help of a 7T MRI scanner (Magnetic Resonator of seven Tesla Varian). The images of the Luffa processed with the OSIRIX® and AMIRA® programs then converted to the STL (STereo Lithography) format for manipulated with CAD / CAM (Computer-Aided Design / Computer-Aided Manufacturing) programs. The results have been successful since it was possible to print by FDM (Fused Deposition Modeling) a scale model of the splint in ABS (Acrylonitrile butadiene styrene), from a module that extracted from the MRI, which tested in a model of a human foot.
Download
Paper Nr: 17
Title:

Left Ventricle Computational Model based on Patients Three-dimensional MRI
Authors:

Maria Narciso, Ana I. Sousa, Fernando Crivellaro, Rui Valente de Almeida, António Ferreira and Pedro Vieira
Abstract: The propagation of electric signal in the heart is bonded to the cardiac muscle (myocardium) geometry and condition. In this regard we aim to build a patient-specific computational model of the myocardium based on noninvasive imaging, with the long-term goal of being used to run simulations in electrophysiology studies. Three-dimensional (3D) Magnetic Resonance Images (MRI) were processed using MATLAB® to build a volumetric mesh which embodies the Left Ventricle (LV) and can later be read by third party applications. This feature was tested with the open source software CHASTE (Cancer, Heart and Soft Tissue Environment) to solve and visualize the propagation of an excitation wave. Furthermore, an algorithm was developed capable of defining the fibre orientation for the resulting mesh, based on the geometry described in literature. This experiment substantiates the expectation that parallel computing simulations of the heart maybe used, in the near future, as a monitoring and diagnostic tool for the assessment of cardiac arrhythmias in clinical practice.
Download
Paper Nr: 34
Title:

Terahertz Reflection Imaging of Paraffin-embedded Human Breast Cancer Samples: Some First Results
Authors:

Mohamed Boutaayamou, Delphine Cerica and Jacques G. Verly
Abstract: Several studies have shown that terahertz (THz) pulsed imaging has the potential of identifying the margins of human breast cancer in paraffin-embedded tissue samples. Before using this technique for the assessment of cancer margins during breast-conserving surgery, it is important to study the validity and reproducibility of previously published results. In the present paper, we describe some first results in the characterization of paraffin-embedded human breast cancer tissue through THz reflection imaging based on measurements provided by a newly acquired THz time-domain spectrometer. First, we measured the THz reflection impulse response of these samples using this spectrometer. Second, we processed, for one selected breast cancer tissue sample, the recorded data to generate preliminary images of (1) several maps of parameters extracted in the time- and frequency-domains, and (2) a map of the absorbance.
Download
Paper Nr: 39
Title:

Object Tracking using CSRT Tracker and RCNN
Authors:

Khurshedjon Farkhodov, Suk-Hwan Lee and Ki-Ryong Kwon
Abstract: Nowadays, Object tracking is one of the trendy and under investigation topic of Computer Vision that challenges with several issues that should be considered while creating tracking systems, such as, visual appearance, occlusions, camera motion, and so on. In several tracking algorithms Convolutional Neural Network (CNN) has been applied to take advantage of its powerfulness in feature extraction that convolutional layers can characterize the object from different perspectives and treat tracking process from misclassification. To overcome these problems, we integrated the Region based CNN (Faster RCNN) pre-trained object detection model that the OpenCV based CSRT (Channel and Spatial Reliability Tracking) tracker has a high chance to identifying objects features, classes and locations as well. Basically, CSRT tracker is C++ implementation of the CSR-DCF (Channel and Spatial Reliability of Discriminative Correlation Filter) tracking algorithm in OpenCV library. Experimental results demonstrated that CSRT tracker presents better tracking outcomes with integration of object detection model, rather than using tracking algorithm or filter itself.
Download