BIOIMAGING 2019 Abstracts


Full Papers
Paper Nr: 4
Title:

Surgical Phase Recognition of Short Video Shots based on Temporal Modeling of Deep Features

Authors:

Constantinos Loukas

Abstract: Recognizing the phases of a laparoscopic surgery (LS) operation form its video constitutes a fundamental step for efficient content representation, indexing and retrieval in surgical video databases. In the literature, most techniques focus on phase segmentation of the entire LS video using hand-crafted visual features, instrument usage signals, and recently convolutional neural networks (CNNs). In this paper we address the problem of phase recognition of short video shots (10s) of the operation, without utilizing information about the preceding/forthcoming video frames, their phase labels or the instruments used. We investigate four state-of-the-art CNN architectures (Alexnet, VGG19, GoogleNet, and ResNet101), for feature extraction via transfer learning. Visual saliency was employed for selecting the most informative region of the image as input to the CNN. Video shot representation was based on two temporal pooling mechanisms. Most importantly, we investigate the role of ‘elapsed time’ (from the beginning of the operation), and we show that inclusion of this feature can increase performance dramatically (69% vs. 75% mean accuracy). Finally, a long short-term memory (LSTM) network was trained for video shot classification, based on the fusion of CNN features and ‘elapsed time’, increasing the accuracy to 86%. Our results highlight the prominent role of visual saliency, long-range temporal recursion and ‘elapsed time’ (a feature ignored so far), for surgical phase recognition.

Paper Nr: 8
Title:

Field of Interest Proposal for Augmented Mitotic Cell Count: Comparison of Two Convolutional Networks

Authors:

Marc Aubreville, Christof A. Bertram, Robert Klopfleisch and Andreas Maier

Abstract: Most tumor grading systems for human as for veterinary histopathology are based upon the absolute count of mitotic figures in a certain reference area of a histology slide. Since time for prognostication is limited in a diagnostic setting, the pathologist will oftentimes almost arbitrarily choose a certain field of interest assumed to have the highest mitotic activity. However, as mitotic figures are commonly very sparse on the slide and often have a patchy distribution, this poses a sampling problem which is known to be able to influence the tumor prognostication. On the other hand, automatic detection of mitotic figures can’t yet be considered reliable enough for clinical application. In order to aid the work of the human expert and at the same time reduce variance in tumor grading, it is beneficial to assess the whole slide image (WSI) for the highest mitotic activity and use this as a reference region for human counting. For this task, we compare two methods for region of interest proposal, both based on convolutional neural networks (CNN). For both approaches, the CNN performs a segmentation of the WSI to assess mitotic activity. The first method performs a segmentation of mitotic cells at the original image resolution, while the second approach performs a segmentation operation at a significantly reduced resolution, cutting down on processing complexity. We evaluate the approach using a dataset of 32 completely annotated whole slide images of canine mast cell tumors, where 22 were used for training of the network and 10 for test. Our results indicate that, while the overall correlation to the ground truth mitotic activity is considerably higher (0.936 vs. 0.829) for the approach based upon the fine resolution network, the field of interest choices are only marginally better. Both approaches propose fields of interest that contain a mitotic count in the upper quartile of respective slides.

Paper Nr: 11
Title:

Quantification of Stromule Frequencies in Microscope Images of Plastids Combining Ridge Detection and Geometric Criteria

Authors:

Birgit Möller and Martin Schattat

Abstract: Plastids are involved in many fundamental biochemical pathways in plants. They can produce tubular membrane out-folds from their surface. These so-called stromules have initially been described over a century ago, but their functional role is still elusive. To identify cellular processes or genetic elements underlying stromule formation screens of large populations of mutant plants or plants under different treatments are carried out and stromule frequencies are extracted. Due to a lack of automatic methods, however, this quantification is usually done manually rendering this step a main bottleneck in stromule research. Here, we present a new approach for quantification of stromule frequencies. Plastids are extracted from microscope images using local wavelet analysis over multiple scales combined with statistical hypothesis testing to resolve competing detections from different scales. Subsequently, for each plastid region evidence for the existence of stromules in its vicinity is investigated applying ridge detection techniques and geometric criteria. Experimental results prove that our approach is suitable to properly identify stromules. Even in microscopy images with a high noise level and distracting signals extracted stromule counts are comparable to those of biological experts.

Paper Nr: 19
Title:

Level Set Segmentation of Retinal OCT Images

Authors:

Bashir I. Dodo, Yongmin Li, XiaoHui Liu and Muhammad I. Dodo

Abstract: Optical coherence tomography (OCT) yields high-resolution images of the retina. Reliable identification of the retinal layers is necessary for the extraction of clinically useful information used for tracking the progress of medication and diagnosis of various ocular diseases. Many automatic methods have been proposed to aid with the analysis of retinal layers, mainly, due to the complexity of retinal structures, the cumbersomeness of manual segmentation and variation from one specialist to the other. However, a common drawback suffered by existing methods is the challenge of dealing with image artefacts and inhomogeneity in pathological structures. In this paper, we embed prior knowledge of the retinal architecture derived from the gradient information, into the level set method to segment seven (7) layers of the retina. Mainly, we start by establishing the region of interest (ROI).The gradient edges obtained from the ROI are used to initialise curves for the layers, and the layer topology is used in constraining the evolution process towards the actual layer boundaries based on image forces. Experimental results show our method obtains curves that are close to the manual layers labelled by experts.

Paper Nr: 20
Title:

Shape Recognition in High-level Image Representations: Data Preparation and Framework of Recognition Method

Authors:

Jagoda Lazarek and Piotr S. Szczepaniak

Abstract: The automatic shape recognition is an important task in various image processing applications, including medical problems. Choosing the right image representation is key to the recognition process. In the paper, we focused on high-level image representation (using line segments), thanks to which the amount of data necessary for processing in subsequent stages is significantly reduced. We present the framework of recognition method with the use of graph grammars.

Short Papers
Paper Nr: 1
Title:

3D Image Deblur using Point Spread Function Modelling for Optical Projection Tomography

Authors:

Xiaoqin Tang, Gerda M. Lamers and Fons J. Verbeek

Abstract: Optical projection tomography (OPT) is widely used to produce 3D image for specimens of size between 1mm and 10mm. However, to image large specimens a large depth of field is needed, which normally results in blur in imaging process, i.e. compromises the image quality or resolution. Yet, it is important to obtain the best possible quality of 3D image from OPT, thus deblurring the image is of significance. In this paper we first model the point spread function along optical axis which varies at different depths in OPT imaging system. The magnification is taken into account in the point spread function modelling. Afterward, deconvolution in the coronal plane based on the modelled point spread function is implemented for the image deblur. Experiments with the proposed approach based on 25 3D images including 4 categories of samples, indicate the effectiveness of quality improvement assessed by image blur measures in both spatial and frequency domain.

Paper Nr: 2
Title:

Models of Learning to Classify X-ray Images for the Detection of Pneumonia using Neural Networks

Authors:

A. A. Saraiva, D. S. Santos, Nator C. Costa, Jose M. Sousa, N. F. Ferreira, Antonio Valente and Salviano Soares

Abstract: This article describes a comparison of two neural networks, the multilayer perceptron and Neural Network, for the detection and classification of pneumonia. The database used was the Chest-X-Ray data set provided by (Kermany et al., 2018) with a total of 5840 images, with two classes, normal and with pneumonia. to validate the models used, cross-validation of k-fold was used. The classification models were efficient, resulting in an average accuracy of 92.16% with the Multilayer Perceptron and 94.40% with the Convolution Neural Network.

Paper Nr: 6
Title:

Enhanced Deep Learning for Pathology Image Classification: A Knowledge Transfer based Stepwise Fine-tuning Scheme

Authors:

Jia Qu, Nobuyuki Hiruta, Kensuke Terai, Hirokazu Nosato, Masahiro Murakawa and Hidenori Sakanashi

Abstract: Deep learning using Convolutional Neural Networks (CNN) has been demonstrated unprecedentedly powerful for image classification. Subsequently, computer-aided diagnosis (CAD) for pathology image has been largely facilitated due to the deep learning related approaches. However, because of extremely high cost of pathologist's professional work, the lack of well annotated pathological image data to train deep neural networks is currently a big problem. Aiming at further improving the performance of deep neural networks and alleviating the lack of annotated pathology data, we propose a full-automatic knowledge transferring based stepwise fine-tuning scheme to make deep neural networks follow pathologist’s perception manner and understand pathology step by step. To realize this conception, we also introduce a new type of target correlation intermediate dataset which can be yielded by using fully automated processing. By extracting rough but stain-robust pathology-related information from unannotated pathology images with handcrafted features, and making use of these materials to intermediately train deep neural networks, deep neural networks are expected to acquire fundamental pathological knowledge in advance so that boosted in the final task. In experiments, we validate the new scheme on several well-known deep neural networks. Correspondingly, the results present solid evidence for the effectiveness and suggest feasibility for other tasks.

Paper Nr: 16
Title:

Image Quality Comparison between Digital and Synthetic 2D Mammograms: A Qualitative and Quantitative Phantom Study

Authors:

P. Barca, R. Lamastra, D. Caramella, A. C. Traino, R. M. Tucciariello and M. E. Fantacci

Abstract: The recent introduction of digital breast tomosynthesis (DBT) have lead to improvements in sensitivity and specificity of breast cancer detection, especially in cases of tumors developed in dense breasts. Since DBT provides tomographic slices of an entire tissue volume, it reduces the inherent tissue overlapping limitation of digital mammography (DM). In addition, DBT combined with DM has been proven to decrease recall and increase invasive cancer detection rates in breast cancer screening. However, the employment of DBT+DM implies a not negligible increment of patients absorbed dose. Therefore, Synthesized mammograms (SMs) generated from the DBT data have been recently introduced to eliminate the need of an additional DM. However, several studies showed differences between DM and SM images and some studies found contrasting results in terms of image quality when DM and SM images were compared. In our phantom study, we objectively compare image quality of SM and DM images in terms of noise, spatial resolution and contrast properties. Additionally, a qualitative analysis of the ACR mammographic phantom was performed in both modalities to assess the detectability of different features. SM images were characterized by different texture with respect to DM images, showing lower overall performances in terms of contrast-to-noise ratio and modulation transfer function. However, the goal of SM images is to provide a useful two-dimensional guide complementary to the DBT dataset and the performances in terms of high-contrast features detectability were satisfactory in comparison to those obtained in DM.

Paper Nr: 17
Title:

Radiotherapy Support Tools, the Brazilian Project: SIPRAD

Authors:

Diego Fiori de Carvalho, José C. Guerrero, Luis M. Zapata, Andrey M. Uscamayta, Heleno C. Vale, Leandro F. Borges, Alexandre C. Bruno and Harley Francisco de Oliveira

Abstract: The radiotherapy planning process (teletherapy) is initially performed by the acquisition of Computed Tomography images of the areas of interest to guide a series of health professionals in the work of vector design of regions of interest for protection (risk organs) and radiation (tumors). All these steps are performed using computational tools that extrapolate measurements and scales in the treatment plan. The efficiency of the treatment depends on the recreation of the patient's positioning on the linear accelerator stretcher with the previously acquired tomography images. For this, in this article, we present three modules of the SIPRAD (Information Systems for Radiation Therapy Planning) project. With the name of Radiotherapy Portal it is able to perform a fusion of planar images of the target region, made on the day of treatment, with the digital recreation (DDR - Digital Reconstructed Radiographs) of this radiograph generated from the Tomography of treatment planning, aiming to improve the reproducibility of the positioning that the radiation dose delivered during all the radiotherapy treatment. The second module named by LYRIA PACS RT provides a client/server architecture for storing, distributing and displaying images from any systems using the DICOM RT Struct, Image, Plan and Dose modes. The third module called Contouring is responsible for the training of new radiotherapists.

Paper Nr: 18
Title:

microCT for Systematic Mouse Phenotyping

Authors:

Frantisek Spoutil, Michaela Prochazkova, Tereza Michalcikova, Ivana Uramova, Sarah Clewell, Vendula Novosadova and Jan Prochazka

Abstract: Phenotyping of mouse mutants is one of the crucial methods for uncovering genetic network at the level of a whole organism which could help us to understand origin of rare diseases, developmental malformations, but also the process of mammalian evolution. For studying morphological aspects of either embryos or adults, the X-ray computed microtomography (microCT) has become a gold standard within the last years. The three-dimensional (3D) context, availability of data to additional analysis (e.g. volumetric, bone density, or body composition), and in-vivo approaches in the case of adults are the main advantages when compared to classic histology and bone morphology. On the other hand, the amount of data is enormous making the data storage and analysis the bottle-neck of the microCT method. To overcome this obstacle, we cooperate with bioinformatics experts to set up automation of the process at maximal possible level. Nevertheless, knowledge and experience of a specialist remain indispensable.

Paper Nr: 5
Title:

Image Segmentation using Gradient-based Histogram Thresholding for Skin Lesion Delineation

Authors:

Pedro M. Pereira, Luis N. Tavora, Rui Fonseca-Pinto, Rui P. Paiva, Pedro A. Assuncao and Sergio M. M. de Faria

Abstract: Image segmentation is a key stage in medical image processing algorithms and machine learning classifiers where identification of discriminative features are of utmost importance. In the case of skin lesions, most of the existing image segmentation approaches aim at minimising some error metric between computed and ground-truth regions of interest (ROI) defined by medical experts, where ROI delineation is not always considered. This paper proposes an image segmentation method for skin lesion delineation, which expands traditional histogram and clustering-based approaches to achieve the best trade-off between both. The proposed method is capable of providing accurate details of the skin lesion borders, without deviating from the coarser borders of the available ground-truth.

Paper Nr: 9
Title:

BOLD Signal Change during Driving with Addition Task using fMRI

Authors:

Ji-Hun Jo, Hyung-Sik Kim, Soon-Cheol Chung and Mi-Hyun Choi

Abstract: This paper uses a driving wheel and pedal (working as an accelerator, brake) equipped with an MR-compatible driving simulator at a speed of 80 km/h when driving and when driving while performing secondary tasks in order to observe differences in neuronal activation (BOLD signal change). The experiments consisted of three blocks, each block consisting of both a Control phase (1 min.) and a Driving phase (2 min.). During the Control phase, the drivers were instructed to look at the stop screen and not to perform driving tasks. During the Driving phase, the drivers either drove or drove while performing addition tasks at 80 km/h. The intensity of activated voxels increased in the addition task condition compared to the driving condition in insula.

Paper Nr: 10
Title:

Polyp Shape Recovery using Vascular Border from Single Colonoscopy Image

Authors:

Hiroyasu Usami, Yuji Iwahori, M. K. Bhuyan, Aili Wang, Naotaka Ogasawara and Kunio Kasugai

Abstract: The shape and size of a colonic polyp is a biomarker that correlates with its risk of malignancy and guides its clinical management. It is the most accurate method for detecting polyps of all sizes, and it allows biopsy of lesions and resection of most polyps, and it is considered nowadays as the gold standard for colon screening. However, there are still open challenges to overcome, such as the reduction of the missing rate. Colonoscopy images usually consist of nonrigid objects such as a polyp, and no approaches have been proposed to recovery shape and absolute size from a single image. Hence, it is a challenging topic to reconstruct polyp shape using computer vision technique. This paper proposes a polyp shape retrieval method based on Shape from Shading (SFS), and this research contributes to mitigating constraint for applying SFS to the single colonoscopy image using vascular border information. Experiments confirmed that the proposed method recovered approximate polyp shapes.

Paper Nr: 12
Title:

Classification of Images of Childhood Pneumonia using Convolutional Neural Networks

Authors:

A. A. Saraiva, N. F. Ferreira, Luciano Lopes de Sousa, Nator C. Costa, José M. Sousa, D. S. Santos, Antonio Valente and Salviano Soares

Abstract: In this paper we describe a comparative classification of Pneumonia using Convolution Neural Network. The database used was the dataset Labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images for Classification made available by (Kermany, 2018) with a total of 5863 images, with 2 classes: normal and pneumonia. To evaluate the generalization capacity of the models, cross-validation of k-fold was used. The classification models proved to be efficient compared to the work of (Kermany et al., 2018) which obtained 92.8 % and the present work had an average accuracy of 95.30 %.

Paper Nr: 14
Title:

Deep Analysis of CNN Settings for New Cancer Whole-slide Histological Images Segmentation: The Case of Small Training Sets

Authors:

Sonia Mejbri, Camille Franchet, Reshma Ismat-Ara, Josiane Mothe, Pierre Brousset and Emmanuel Faure

Abstract: Accurate analysis and interpretation of stained biopsy images is a crucial step in the cancer diagnostic routine which is mainly done manually by expert pathologists. The recent progress of digital pathology gives us a challenging opportunity to automatically process these complex image data in order to retrieve essential information and to study tissue elements and structures. This paper addresses the task of tissue-level segmentation in intermediate resolution of histopathological breast cancer images. Firstly, we present a new medical dataset we developed which is composed of hematoxylin and eosin stained whole-slide images wherein all 7 tissues were labeled by hand and validated by expert pathologist. Then, with this unique dataset, we proposed an automatic end-to-end framework using deep neural network for tissue-level segmentation. Moreover, we provide a deep analysis of the framework settings that can be used in similar task by the scientific community.

Paper Nr: 21
Title:

Separation of Foreign Patterns from Native Ones: Active Contour based Mechanism

Authors:

Piotr S. Szczepaniak

Abstract: This position paper presents an approach to the problem of foreign pattern rejection from a data set containing both native and foreign patterns. On the one hand, the approach may be regarded as classic in the sense that it is based on well-known concepts: class-contra-all-other classes or class-contra-class, but on the other hand, the novelty lies in the (embedding) application of potential active hypercontour, which is a powerful method for solving classification problems and may be applied as a binary or multiclass classifier.

Paper Nr: 22
Title:

Segmentation of Shoulder MRI Data for Musculoskeletal Model Adaptation

Authors:

Tomáš Ryba and Zdeněk Krňoul

Abstract: Applying image processing techniques to medical images has already brought many useful applications. This work is focused on using these methods in the process of adapting a musculoskeletal model of the shoulder joint. Comparing the model of healthy individuals and the patient with joint damage leads to a subject-specific convalescence treatment. This work describes the procedure for segmentation of magnetic resonance imaging (MRI) of the shoulder joint. Firstly three bones inside the shoulder area: humerus, clavicle, scapula are identified and thereby it provides initial reference objects. A major step is the segmentation of the deltoid muscle needed for the subsequent adaptation of the musculoskeletal model. This step is challenging in terms of image processing due to the closeness of soft tissues, which are almost identical in intensity and the boundaries between them are often barely visible. The approach to resolving this problem is described and possible improvements and future work are described.

Paper Nr: 24
Title:

Biomedical Device for Early Breast Cancer Detection: Device Performance Improving by Plasmonic-Photonic Mask

Authors:

Sanem Meral, Ezel Yalcinkaya, Metin Eroglu, Ahmad Salmanogli and H. S. Gecim

Abstract: In this article, a new device to detect breast cancer at an early stage, is presented. The main advantages of the device are its easy operational procedure, portability, high accuracy due to usage of plasmonic-photonic mask and the low cost. In fact, the novelty of the device presented is to apply the new mask called plasmonic-photonic mask for precise analysis of the captured images. In the early stage of the work, a phantom model is employed and the operation of the system is realized. It is shown that the image processing toolbox is safely matched with the device. It should be noted that for the in-vivo imaging, the device should be completed and equipped with a high accuracy charge coupled device (CCD) and laser.