CBA Seminars in Fall 2022
| Increasing Trust in Alzheimer's Disease Forecasting
Alzheimer’s disease (AD) is the most common form of dementia, causing progressive cognitive impairment, disorientation, and memory loss. Despite years of clinical trials, there is currently no cure for AD and, by the time of diagnosis, as much as 60% of brain matter is already lost. Thus, forecasting AD years ahead of onset is critical for attempts at early treatment, the selection of subjects for clinical trials, and to facilitate neurologists’ study of the disease. Past research on machine learning for AD prediction was limited to using cognitive test scores and highly engineered volumetric features, while failing to leverage the potential information found in brain MRIs. In this project, we study how RNN and CNN can contribute to a mutual information based unsupervised encoder and extract latent features from 3D MRIs. The future goal of the project lies in forecasting the disease on a longer time gap without loss of accuracy and robustness.
|Madalina Fiterau, University of Massachusetts Amherst, USA
| Oral Cancer Screening by Detection of Malignancy Associated Changes in Whole Slide Oral Cytology Images using Deep Convolutional Neural Network and Vision Transformer Based Frameworks
Cases of Oral Cancer are increasing around the world. Oral Squamous Cell Carcinomas constitute the majority of all Oral Cancer cases and arise from the oral epithelium. Although this type of Oral Cancer is highly accessible to clinicians as they are superficial, they are often discovered late. To improve early detection in order to increase the chances of survival, we explore Deep Convolutional Neural Network and Vision Transformer based frameworks on Whole Slide Oral Cytology Images to detect Malignancy Associated Changes (MACs). MACs are subtle changes that happen to the normal cells which are in the close vicinity of a tumor. These changes may be in size, shape, and chromatin structure of the nucleus. The dataset used for the ongoing work has 12 Whole Slide Images (WSIs) of Liquid Based Cytology (LBC) samples (6 Normal and 6 Malignant) collected from 12 patients. We propose a three stage framework : i) Nuclei detection using a Fully Convolutional Regression Network (FCRN), ii) Selection of the best focal level from the z-stack for each of the detected nuclei, iii) Classification of the extracted patches (using standard Deep CNN and Vision Transformer architectures) into pathological conditions: Normal and Cancerous on the basis of absence or presence of MACs.
| Label free species classification of bacteria growing in microfluidic chip imaged by phase-contrast microscopy
We classify time-lapses or "video-clips" of the growing bacteria using video-clip and image classification networks. A prototype of an active-learning data labeling tool to perform object detection of bacteria using fluorescence stains will be demonstrated.
| Building Convolutional Feature Extractors from Image Markers
The success of a Convolutional Neural Network (CNN) mainly depends on its feature extractor (encoder). For example, an encoder that successfully maps samples from distinct classes into separated subspaces of its output feature space can reduce the classifier's depth into a single decision layer. However, training deep models with backpropagation from scratch requires considerable human effort in data annotation and hyperparameter adjustments, leaving unanswered questions, such as: How many annotated samples are required to train the model? What is the simplest model to solve the problem? Can we build the model layer by layer and explain its decisions? We have addressed such questions by combining semi-automated data annotation, information (data) visualization, and model construction layer by layer. This talk focus on the latter topic by presenting FLIM (Feature Learning from Image Markers). FLIM builds encoders from strokes drawn by the user on relevant regions of very few training images per class. By that, we aim to considerably reduce human effort in data annotation with more user control and understanding of the model. FLIM estimates the kernels of each convolutional layer from patches centered at the marked pixels and their activations in the previous layer. The process relies on patch clustering rather than backpropagation to select the most representative patches to activate similar regions in new images. The resulting CNN usually outperforms the same architecture trained from scratch by backpropagation. The talk presents results compared to deep models for image classification and segmentation problems and discusses the open problems in FLIM to motivate collaboration.
|Alexandre X. Falcao, University of Campinas, Brazil
|InfraVis course: Visualizing data of varying dimensions
|InfraVis course: Visualization of geographic data
| Improving the training of deep learning with equivariant neural networks
While ordinary CNN classifiers are invariant to translations of the inputs, equivariant neural networks extend this property to other symmetries, such as rotations. We demonstrate these methods in biomedical image analysis, e.g., virus classification in transmission electron microscopy images. The equivariant neural networks typically show improvements over baseline CNNs in terms of training convergence time, overfitting, or the amount of training data needed.
|Karl Bengtsson Bernander
| Image and data analysis for spatial transcriptomics and tissue architecture
This half-time seminar will be about all the projects I have worked on during my PhD education. The topics included will be: spatial statistics, spatial analysis of bladder cancer samples, the TissUUmaps project, spatial analysis of brain tissue micro-arrays, graph clustering method developed for ISS data, and decoding of in situ one fluorescent cycle microscopy data.
External reviewer: Åsa Björklund, WABI
| Applying Handwritten Text Recognition to Astrid Lindgren's Stenographic Manuscripts
Swedish children's book author Astrid Lindgren has written many of her drafts and revisions in the Swedish "Melin" system of shorthand. Following an extensive crowdsourcing effort over the last two years, a portion of her stenographed records have been transliterated. Based on this data, various HTR models are currently being trained and evaluated to determine their applicability to the remaining, untransliterated manuscripts in the Lindgren collection.
In this seminar, I will briefly introduce our new dataset, followed by a presentation of the latest HTR results.
Universitetshuset, sal IV
| Professor installation lecture (in Swedish):
Se och lär – men förklara också
Universitetshuset, sal IV
| Professor installation lecture (in Swedish):
Bildanalys och självkörande mikroskop i jakten på virus och annat livsviktigt smått
| Contrastive Multi-modal Image Representations for Cross-modal Image Retrieval
Contrastive learning can be used to learn to learn multi-modal image representation, embedding images of multiple modalities into a shared space. Such contrastive multi-modal image representations (called CoMIRs) have been developed and successfully used for the task of cross-modal image registration. In this seminar, we will discuss their applicability in the scope of cross-modal image retrieval. Experimenting with a multi-modal brightfield microscopy (BF) and second harmonic generation (SHG) dataset where different modalities share few common structures, we show that reducing a cross-modal image retrieval problem to a mono-modal one through CoMIR representations positively affects the retrieval success. We compare cross-modality retrieval success on raw images, CoMIR embeddings, CycleGAN- and Pix2Pix-generated fakes and discuss the possible reasons behind the consistently superior success of the CoMIR based retrieval over the others.
| Explainable Deep Learning Methods for Human-Human and Human-Robot Interaction
This half-time seminar is an overview of the research I have done in the first two years of my PhD. We will discuss training video CNNs for affect recognition, the challenges of applying end-to-end learning to social interaction datasets, and initial results in using human-likeness to compare XAI techniques.
External reviewer: Albert Ali Salah, Utrecht University, The Netherlands
|Marc Fraile Fàbrega
| Why zebrafish and what can we use it for?
Medical research is dependent on animal models to study the cause of human diseases, as well as the development of new therapies. The rodent is the most widely used research model. However, during recent decades the use of the zebrafish (Danio rerio) as a model organism has increased significantly. This small fish shares many similarities with humans and can easily be genetically modified to generate models of a variety of human diseases. The fish is easy to maintain, breed and has a transparent body, all that makes them well suited for research. There are many groups at Uppsala University (UU) that currently use the zebrafish as a model organism. However, to utilize the full potential of the research, analysis methods must be developed to quantify and compare phenotypes. At UU, we have together with several groups developed and established different acquisition and analysis pipelines. This talk will present results from the ongoing zebrafish research here at UU.
| Can we improve representation learning for multimodal image registration by additional supervision on intermediate layers?
Multimodal imaging is important in digital pathology, typically requiring the alignment of images. Contrastive learning can be used to learn representations of multimodal images to reduce the challenging task of multimodal image registration to a monomodal one. Previously, additional supervision on intermediate layers in contrastive learning has improved image classification performance on biomedical datasets. In this study we evaluate if a similar approach can boost the registration performance. We explore three different approaches to add additional supervision to the latent representations of multimodal images and evaluate the use of three different similarity/distance measures (Mean Squared Error, Cosine Similarity, L1 norm) on these latent features for each approach. Our results show that representations learnt with no additional supervision on latent features perform best in the downstream task of registration on two publicly available biomedical datasets which are highly relevant in cancer diagnostics and research.
| Leveraging network representations for cancer histopathology and spatial-omics
This half-time seminar will focus on Eduard's work during his PhD time. He has been exploring how deep network representations of hematoxylin and eosin (H&E) slides describe tissue morphology. Some of the work includes correlating them with spatial gene expression and using them for interactive annotation. Work in progress also includes using representations on other microscopy images and using them in a weakly supervised framework.
External reviewer: Kevin Smith, KTH
|Eduard Chelebian Kocharyan