148 research outputs found
Signal Processing for Newborn Survival : from labour to resuscitation
No full textPhD thesis in Information technologyStillbirths are a worldwide challenge, with an estimated 2.6 million stillbirths in 2015, of these 1.3 million are estimated to have died during labour and birth, i.e. fresh stillbirth. In addition to the 2.6 million, one million newborns die within their first and only day of life. Complications due to birth asphyxia are the primary cause of these deaths. The vast majority, 98%, of stillbirths and early neonatal deaths are found in low resource settings.
This thesis investigates two main challenges related to neonatal deaths, 1) fetal heart rate (FHR) and labour analysis, and 2) improving newborn resuscitation. The FHR is known to be important for effectively assessing the well-being of the fetus during labour. In high resource countries, the FHR is measured using cardiotocography for all high-risk labours. While in low income countries, assessment of the FHR is often done manually using a Pinard. With the use of continuous FHR monitoring in low income countries, abnormalities in FHR could potentially be identified at an earlier stage. In this thesis, we facilitate for further analysis of FHR signals by proposing a method to remove less trustworthy time periods of the measured signal, such as noise. And how missing data can be estimated using dictionary learning to allow for continuous analysis. The FHR signals should be interpreted in combination with the uterine activity. We therefore propose a method for detecting uterine contractions using an accelerometer mounted together with the Doppler ultrasound FHR sensor. Finally, we explore how FHR develops during labour and how this trend differs for labours with a normal and adverse outcomes.
For newborns who are unable to start breathing themselves after birth, immediate help from the healthcare workers are crucial. We therefore explore which parameters during newborn resuscitation are important for the resuscitation outcome. One of the identified parameters is the amount of stimulation, i.e. rubbing the back of the newborn. To get a greater understanding of how stimulation affects newborn resuscitation, and how it should be applied, a large number of annotated resuscitation episodes are required. Manual annotation is both a time consuming and challenging process for the reviewer. We have proposed a complete system for automatically annotating stimulation by using ECG and accelerometer signals measured on the abdomen of the newborn
Deep Learning-Driven Diagnostic and Prognostic Solutions for Histopathological Images of Bladder Cancer
No full textPhD thesis in Information technologyThis thesis presents a comprehensive investigation into the development and application of advanced computational techniques for the extraction of crucial diagnostic and prognostic information from histological images of non-muscle invasive bladder cancer (NMIBC). Computational pathology (CPATH) relies on digitized high-resolution tissue samples, referred to as whole slide images (WSIs). Histological examination of WSIs plays a pivotal role in the diagnosis and prognosis of NMIBC. The primary focus of this research is the utilization of deep learning algorithms to automatically analyze histological images and extract visual cues with diagnostic and prognostic significance.
With respect to diagnostics, several convolutional neural network architectures are designed and trained on diverse datasets of NMIBC tissue specimens to identify and classify key histological features, including tumor grading and staging. Moreover, the variability of histological visual features between pathology laboratories during the training of convolutional neural network (CNN) models is questioned. Emphasis is placed on the development of label-efficient guidelines for domain-adapting deep learning models. In addition, an architecture for machine learning is introduced to stratify regions of interest (ROIs) in weakly supervised learning. This additional data stratification aids in localizing ROIs and mitigating cross-noise variability among them.
Furthermore, this thesis explores the integration of deep learning techniques for prognostic assessment. Through an analysis of the relative spatial distribution among urothelium and contingent stromal immune cells, our model predicts patient treatment outcomes and the likelihood of recurrence with a high degree of precision. These prognostic models provide invaluable support to clinicians in customizing personalized treatment strategies and offering patient counseling.
The work presented in this thesis represents a substantial advancement toward improving the diagnostic and prognostic capabilities in the management of NMIBC. Leveraging the potential of computational analysis, we offer pathologists state-of-the-art tools to augment diagnostic precision and optimize patient care, ultimately contributing to better outcomes and quality of life for individuals affected by this prevalent form of bladder cancer
Analysis of Thermal and Visible Light Video Data from Birth Episodes Using Artificial Intelligence
No full textPhD thesis in Information technologyGlobally, 10% of newborns experience insufficient breathing at birth and require immediate assistance to achieve cardiopulmonary stability. Newborn resuscitation is time-critical, and providing ventilation within the “golden minute”–the first minute after birth–significantly lowers the risk of death and long-term complications. Analyzing newborn resuscitation videos has proven valuable for evaluation, debriefing, and training. However, accurately assessing guideline effectiveness and optimizing the timing of Newborn Resuscitation Algorithm Activities (NRAA) requires a large dataset of precisely annotated episodes. Manual annotation is time-consuming, inefficient, and raises privacy concerns. Additionally, videos recorded from resuscitation episodes lack crucial information on the Time of Birth (ToB) and transfer duration to the resuscitation station. To properly evaluate the treatment given, NRAA timelines must be relative to an accurate ToB, recorded with second precision. Currently, clinical ToB recording methods often rely on manual processes with minute precision, making them prone to inaccuracies.
This thesis investigates deep learning algorithms for automating the generation of detailed NRAA timelines, including ToB and newborn resuscitation activities. The conducted work focuses on three key areas. First, we explore video-based methods to automatically recognize overlapping resuscitative interventions and generate event timelines from newborn resuscitation episodes. Second, we delve into the combination of deep learning methods and thermal imaging to detect the ToB, addressing the challenges of using infrared technology and refining ToB estimations through signal processing techniques. Finally, we introduce explainability methods in our application to shed light on the decision-making process, enhancing trust and transparency.
The contributions presented in this thesis represent a significant advancement toward improving neonatal care. Moreover, to the best of our knowledge, no prior work has addressed automated ToB detection. Our findings offer a potential solution for evidence-based research, quality improvement, clinical documentation, debriefing, and real-time decision support in neonatology
Automated Grading of Bladder Cancer using Deep Learning
No full textUrothelial carcinoma is the most common type of bladder cancer and is among the cancer types with the highest recurrence rate and lifetime treatment cost per patient. Diagnosed patients are stratified into risk groups, mainly based on the histological grade and stage. However, it is well known that correct grading of bladder cancer suffers from intra- and interobserver variability and inconsistent reproducibility between pathologists, potentially leading to under- or overtreatment of the patients. The economic burden, unnecessary patient suffering, and additional load on the health care system illustrate the importance of developing new tools to aid pathologists.
With the introduction of digital pathology, large amounts of data have been made available in the form of digital histological whole-slide images (WSI). However, despite the massive amount of data, annotations for the given data are lacking. Another potential problem is that the tissue samples of urothelial carcinoma contain a mixture of damaged tissue, blood, stroma, muscle, and urothelium, where it is mainly the urothelium tissue that is diagnostically relevant for grading.
A method for tissue segmentation is investigated, where the aim is to segment WSIs into the six tissue classes: urothelium, stroma, muscle, damaged tissue, blood, and background. Several methods based on convolutional neural networks (CNN) for tile-wise classification are proposed. Both single-scale and multiscale models were explored to see if including more magnification levels would improve the performance. Different techniques, such as unsupervised learning, semi-supervised learning, and domain adaptation techniques, are explored to mitigate the challenge of missing large quantities of annotated data.
It is necessary to extract tiles from the WSI since it is intractable to process the entire WSI at full resolution at once. We have proposed a method to parameterize and automate the task of extracting tiles from different scales with a region of interest (ROI) defined at one of the scales. The method is reproducible and easy to describe by reporting the parameters.
A pipeline for automated diagnostic grading is proposed, called TRIgrade. First, the tissue segmentation method is utilized to find the diagnostically relevant urothelium tissue. Then, the parameterized tile extraction method is used to extract tiles from the urothelium regions at three magnification levels from 300 WSIs. The extracted tiles form the training, validation, and test data used to train and test a diagnostic model. The final system outputs a segmented tissue image showing all the tissue regions in the WSI, a WHO grade heatmap indicating low- and high-grade carcinoma regions, and finally, a slide-level WHO grade prediction. The proposed TRIgrade pipeline correctly graded 45 of 50 WSIs, achieving an accuracy of 90%
Automatic AI-Driven segmentation of Acute Ischemic Stroke Regions with CT Perfusion Images
No full textPhD thesis in Information technologyThis thesis investigates artificial intelligence (AI) methodologies to automatically delineate ischemic areas of brain Computed Tomography Perfusion (CTP) scans acquired at hospital admission in patients suspected of acute ischemic stroke (AIS). Stroke, a critical neurological disorder, has a substantial socio-economic impact and a tremendous effect on the quality of life of afflicted subjects. Time is essential for dealing with this neurological disorder: every minute millions of brain cells die during a cerebral stroke. Consequently, developing accurate and rapid automatic prediction techniques for identifying the location and size of ischemic regions, including tissue with an extremely high probability of infarction (core) and potentially recoverable tissue (penumbra), are of serious clinical interest.
CTP is a fast and widely used 4D imaging modality employed upon hospital admission for evaluating stroke severity and aiding in treatment planning. Automated segmentation methods for CTP need to be perform fast and within the golden hour for identifying tissue-at-risk and prepare treatments. Current methods primarily rely on clinically interpretable 3D parametric maps derived from these scans. Parametric maps are also generally adopted by neuroradiologists for assessing this neurological disorder. However, few segmentation analyses have investigated the usage of AI pipelines with 4D CTP as input. These segmentation methods only focus on segmenting already infarcted areas or core regions, neglecting the penumbra. Nonetheless, predicting penumbra areas can hold significant importance in treatment planning.
This thesis explores conventional supervised and semi-supervised AI approaches to segment both penumbra and core areas. Parametric maps and CTP scans have been adopted as input for Machine Learning (ML) and Deep Learning (DL) algorithms to determine the most suitable input data for the automatic segmentation task. Scans from subjects of differentages and severity groups have been leveraged for training the ML and DL models, thus simulating real-world scenarios. Exploiting 4D CTP scans as input provided promising segmentation results on this dataset, regardless of the severity group, but it produced over-segmentation on large ischemic areas. Few-shot learning approaches returned promising outcomes; however, the results are still distant from supervised architectures. The thesis demonstrated the feasibility of employing CTP studies as input modality for segmenting both ischemic regions (penumbra and core) at hospital admission
Automatic segmentation of bone, skin and synovitis in ultrasound images of finger joints
No full textMaster's thesis in Cybernetics and signal processingRheumatoid arthritis (RA) is estimated to affect between 0.3 to 1.5 % of the population. It tends to strike individuals between the ages of 35-50, which is their working age, with every third individual diagnosed with RA becoming work disabled, and up to 85% of the individuals who still can work losing almost 40 days per year on average. Therefore, an accurate measurement of disease activity is crucial to provide adequate treatment and care for patients. The first stage in RA is inflammation of the synovial membrane which is called synovitis. Using ultrasonography has proven to provide useful information regarding the disease activity. The assessment of disease activity has until now been done visually by doctors by grading the synovitis from 0-3 in the ultrasound images. Making a software to automate these assessments in order to reduce the number of human-dependent discrepancies can be advantageous.
Materials given in this thesis came from the Norwegian and Polish collaborative project, MEDUSA. They included ultrasound images of finger joints and manually annotated data which was used for similarity measurement. The objective of this thesis has been to segment the synovitis in the ultrasound images automatically. Since it develops from the joint area towards the skin, it was nec-essary to segment skin and bones first. Multiple image processing techniques were tested for the proposed system for segmentation of bone skin and synovitis. Novel methods for segmentation and location of these features were also developed. All the proposed methods were implemented using MATLAB.
The similarity measurement was done by computing the modified Hausdorff distance for bone and skin, whereas the Dice coefficient was used for comparing the synovitis with the annotation data. The results show that the proposed system for segmentation of bone and skin functioned well with 80% of the segmented bone and skin features having a distance under 20px to the annotation data. However, one of the two bones had only 55% under 20px, but had a median of 11px. The proposed system for segmentation of the synovitis gave an overall low Dice coefficient, with the best result giving a median and mean Dice of 58 and 54 respectively using Region growing. However, when inspecting the images visually, most of the segmented synovitis seemed descent.
It was concluded that even though the skin and bone segmentation was good, the proposed methods for segmentation of the synovitis did not yield satisfactory results for future grading of it
Classification of histological images of bladder cancer using deep learning
No full textMaster's thesis in Cybernetics and signal processingIn Norway bladder cancer is the fourth most common cancer type among men, with an almost 70 % increase in incidence the past four decades. For women, the increase has been about 40 %.
The histological images of bladder cancer are investigated by a pathologist to determine the grade and stage of cancer. In addition, the risk of recurrence and progression are also diagnosed. This is done manually by studying the histological images, but reproducibility of these results are low. To aid the pathologist, a proposed automatic system have been designed in this thesis consisting of six steps. Step one to four have been studied and experimented in detail, and step five and six are considered as future work.
The histological images are divided into smaller tiles, where each tile consists of one of several different categories; cancer tissue, damaged tissue, other tissue, blood or background. The aim is to make a system which automatically separates all tiles containing cancer tissue from the rest, as these have the potential to diagnose the cancer grade, stage, recurrence and progression.
To distinguish the different categories from each other, a classification system was constructed consisting of an autoencoder and a classifier trained in a semi-supervised fashion. The autoencoder was trained on 943,127 unlabeled tiles, extracted from seven histological images. Next, the encoder part of the autoencoder was connected to the classifier which was fine-tuned on 152,312 labeled images.
For evaluating the performance of the classifier, 10-fold cross-validation was calculated. Accuracy of the best classifier on a five class dataset was 97.7 % with a standard deviation of 3.2 %
Image processing on histopathological images of urothelial carcinoma – assessment of immune cells
No full textMaster's thesis in Automation and Signal processingBladder cancer is the 6th most common cancer in the world, where urothelial carcinoma is the most common one. Bladder cancer is one of the most economically expensive cancers to treat, as follow up is needed over a long period of time. Through extensive research, it has been indicated that the amount of tumor infiltrating lymphocytes(TIL) can have a positive impact on the relapse rate in conjunction with treatment.
This paper concentrates on image processing to identify, and analyze the amount of TIL cells in histological images of bladder tissue. The objective of this thesis is to locate all cells in a histological image, and to train a classifier to predict if a cell is a TIL or not. The end goal is to automatically determine the amount of TILs in an image which in turn can be used to predict the effectiveness of cancer treatment. A sub set of microscopic tissue samples has been derived from digitized samples, made available by Stavanger Universitetssykehus, to be able to analyze the quantitative performance of the proposed system.
Using a distance transform, in conjunction with pre-processing methods, to 93% of the cells in the histological images were found. A side effect was that there were wrongly located multiple cell centers for some cells, in addition to other non-cell objects in the histological images. Prediction of the located cells, using histogram features, was able to achieve 92% accuracy. Using local binary pattern features, the prediction accuracy was reduced to 73%. Synthetic over-sampling was introduced as the prediction showed a higher accuracy for correctly predicted non-TILs, but this proved to decrease the quantitative performance
Bildebehandling og dype nevrale nett for deteksjon av immunceller på histologiske bilder av blærekreft
No full textMaster's thesis in Automation and signal processingBladder cancer is the tenth most common cancer type, where urothelial carcinoma is the most common type of bladder cancer. Bladder cancer has been classified as the most expensive type of cancer per patient, as the need for post-treatment monitoring often lasts the rest of the patient’s life. A pathologist needs to diagnose and evaluate the risk of progression and relapse from analyzing histological images.
Recent research shows a correlation between the number of regulatory T-cells and which patients that get progression to a higher cancer grade. Today a computer randomly picks out a sub-set of cells, that is to be manually counted and classified; this will serve as an estimation for regulatory T-cells compared to other cells. This paper proposes a more automated solution to aid in analyzing histological images for the number of regulatory T-cells and other cells present.
The two proposed systems are using classical image processing to find and classify the cells based on color and using a convolutional neural network to detect and classify smaller parts of the images. Both systems will attempt to estimate the number of regulatory T-cells compared to other cells.
The classical image processing had an underestimation of 4.7% for regulatory T-cells while having a 4.5% overestimation of other cells. The convolutional neural network showed a correlation between the number of classifications and the actual amount of cells but requires further work to be usable
Diagnosis, Localization, and Prognosis of Melanoma in WSIs with a Complete Pipeline by Digital Pathology
No full textThe most dangerous and aggressive form of skin cancer is melanoma, responsible for 90% of skin cancer mortality. Early detection of melanoma plays a crucial role in the prognostic outcome. The diagnostic has to be performed by a pathologist, which is time- consuming. The recent increase in melanoma incidents indicates the growing demand for a more efficient diagnostic process.
This thesis’s main objective is to develop a pipeline utilizing two independent pre- trained models built on the VGG16 architecture. This pipeline consists of a diagnostic and a prognostic model. The diagnosis model is responsible for localizing malignant patches in WSIs and giving a patient-level diagnosis. The prognosis model uses the output from the diagnosis model to provide a patient-level prognosis. The complete pipeline provides both a prognostic and a diagnostic tool, which can be used by a pathologist when evaluating Whole Slide Images (WSIs). A total of 243 WSIs were provided by Stavanger University Hospital for this thesis. All have been provided a patient-level label. 203 of the WSIs were used for parameter tuning and 40 were used for testing.
The diagnosis model performed with a 100% accuracy when evaluated on the original test, which was provided together with the training set. The prognosis model also per- formed well on the original dataset, with an accuracy of 0.7885. The model’s capability to predict diagnosis and prognosis decreases significantly when being introduced to the new dataset. In addition to developing the pipeline, some parameters for the diagnosis model was found using a ROC cuve. By using the new parameters for the diagnosis model on the validation set, the performance of the diagnosis model increased when using the test set. The prognosis model performed relatively equally in all experiments. A correlation between the number of patches in a WSI and the number of patches predicted malignant was discovered and counteracted by altering the patient-level threshold calculation method.The most dangerous and aggressive form of skin cancer is melanoma, responsible for 90% of skin cancer mortality. Early detection of melanoma plays a crucial role in the prognostic outcome. The diagnostic has to be performed by a pathologist, which is time- consuming. The recent increase in melanoma incidents indicates the growing demand for a more efficient diagnostic process.
This thesis’s main objective is to develop a pipeline utilizing two independent pre- trained models built on the VGG16 architecture. This pipeline consists of a diagnostic and a prognostic model. The diagnosis model is responsible for localizing malignant patches in WSIs and giving a patient-level diagnosis. The prognosis model uses the output from the diagnosis model to provide a patient-level prognosis. The complete pipeline provides both a prognostic and a diagnostic tool, which can be used by a pathologist when evaluating Whole Slide Images (WSIs). A total of 243 WSIs were provided by Stavanger University Hospital for this thesis. All have been provided a patient-level label. 203 of the WSIs were used for parameter tuning and 40 were used for testing.
The diagnosis model performed with a 100% accuracy when evaluated on the original test, which was provided together with the training set. The prognosis model also per- formed well on the original dataset, with an accuracy of 0.7885. The model’s capability to predict diagnosis and prognosis decreases significantly when being introduced to the new dataset. In addition to developing the pipeline, some parameters for the diagnosis model was found using a ROC cuve. By using the new parameters for the diagnosis model on the validation set, the performance of the diagnosis model increased when using the test set. The prognosis model performed relatively equally in all experiments. A correlation between the number of patches in a WSI and the number of patches predicted malignant was discovered and counteracted by altering the patient-level threshold calculation method
- …
