contrastive learning

Image model embeddings for digital pathology and drug development via self-supervised learning

This paper introduces DIME, a self-supervised learning pipeline for training image patch embeddings from whole slide images, enabling effective downstream tasks with limited data.

MM-Deacon: Multimodal molecular domain embedding analysis via contrastive learning

MM-Deacon generates multimodal molecular embeddings using contrastive learning to align SMILES and IUPAC representations, improving molecule clustering and cross-modal search.

Examining Batch Effect in Histopathology as a Distributionally Robust Optimization Problem

This paper shows that using abstention in a patch-based classifier improves performance in histopathology tasks while reducing reliance on spurious features.

DeepDDS: deep graph neural network with attention mechanism to predict synergistic drug combinations

DeepDDS uses a graph neural network with attention to predict synergistic drug combinations, outperforming existing methods on benchmark and independent test data.

Inexpensive monitoring of flying insect activity and abundance using wildlife cameras

This paper introduces a low-cost method using wildlife cameras to monitor flying insect activity and abundance, demonstrating its effectiveness for Bogong moths in remote areas.

CoSTA: Unsupervised Convolutional Neural Network Learning for Spatial Transcriptomics Analysis

CoSTA uses a convolutional neural network to analyze spatial transcriptomics data, identifying gene relationships based on spatial patterns rather than pixel-level correlation.

A Self-Supervised Deep Neural Network for Image Completion Resembles Early Visual Cortex fMRI Activity Patterns for Occluded Scenes

This paper compares a self-supervised CNN for image completion with fMRI data from early visual cortex, finding the CNN's decoder pathway more similar to brain activity than a supervised model.

MICRA-Net: MICRoscopy Analysis Neural Network to solve detection, classification, and segmentation from a single simple auxiliary task

MICRA-Net is a weakly-supervised neural network that uses a simple classification task to perform segmentation, detection, and enumeration on microscopy images with limited annotated data.

Deep learning assessment of cultural ecosystem services from social media images

The paper uses deep learning to automate the assessment of cultural ecosystem services from social media images, showing good performance with pre-trained models.

Beyond category-supervision: instance-level contrastive learning models predict human visual system responses to objects

This paper introduces a self-supervised model that learns instance-level representations of images, achieving performance comparable to category-supervised models in predicting human ventral visual stream responses.

Noise Reconstruction & Removal Network: A New Architecture to Denoise FIB-SEM Images

The paper introduces an unsupervised Noise Reconstruction and Removal Network for denoising FIB-SEM images, achieving results comparable to or better than supervised methods.

Using a two-stage convolutional neural network to rapidly identify tiny herbivorous beetles in the field

The paper develops a two-stage CNN method to improve identification of tiny beetles in the field, achieving higher precision than single-stage models.

Predicting phenotypes from genetic, environment, management, and historical data using CNNs

This paper uses CNNs to predict phenotypes from genetic, environmental, and management data, showing improved accuracy with pre-training on historical data.

Uncovering Interpretable Fine-Grained Phenotypes of Subcellular Dynamics through Unsupervised Self-Training of Deep Neural Networks

This paper introduces DeepHACX, a self-training deep learning framework for interpretable fine-grained phenotyping of live cell dynamics, revealing specific responses to pharmacological perturbations in migrating epithelial cells.

Sampling from Disentangled Representations of Single-Cell Data Using Generative Adversarial Networks

The paper develops MichiGAN, a generative model combining VAE and GAN strengths, to sample from disentangled single-cell RNA-seq representations and predict drug responses.

Protein contact map prediction using multiple sequence alignment dropout and consistency learning for sequences with fewer homologs

The paper improves protein contact map prediction for sequences with few homologs by using consistency learning on reduced multiple sequence alignments.

An Explainable Deep Learning Approach for Multimodal Electrophysiology Classification

The paper introduces a gradient-based explainability method for multimodal electrophysiology classifiers, showing EEG's dominance in sleep stage classification.

Innocent until proven guilty: Privacy-preserving search over a central CODIS criminal database from the field

The paper introduces a cryptographic algorithm that allows private DNA profile matching against a CODIS database without revealing the profile to the database provider.

Translation of cellular protein localization by generative adversarial network

This study uses a generative adversarial network to translate actin filament images into protein localization images, showing high similarity for actin-related proteins but lower for microtubules.

Efficient and scalable integration of single-cell data using domain-adversarial and variational approximation

DAVAE integrates single-cell data across samples, technologies, and modalities by removing batch effects and enabling transfer learning and cell type prediction.

Unsupervised temporal consistency improvement for microscopy video segmentation with Siamese networks

The paper improves microscopy video segmentation by retraining a CNN with a Siamese network setup using unlabelled video frames.

Improving the sensitivity of cluster-based statistics for fMRI data

This paper improves the sensitivity of cluster-based statistics for fMRI data using novel cluster definitions and a new method to combine test statistics, without sacrificing spatial specificity.

Network Analysis and Transcriptome Profiling Identify Autophagic and Mitochondrial Dysfunctions in SARS-CoV-2 Infection

The paper identifies autophagic and mitochondrial dysfunctions in SARS-CoV-2 infection through transcriptome profiling and network analysis of cell lines and human samples.

Explaining decisions of Graph Convolutional Neural Networks: patient-specific molecular subnetworks responsible for metastasis prediction in breast cancer

The paper introduces GLRP, a method to explain Graph-CNN decisions in breast cancer metastasis prediction, identifying patient-specific molecular subnetworks that align with clinical knowledge.

Unsupervised content-preserving transformation for optical microscopy

This paper introduces an unsupervised image transformation method for optical microscopy that avoids data annotation and achieves high-fidelity image transformations across modalities.

Transfer learning framework for cell segmentation with incorporation of geometric features

The paper introduces a transfer learning framework for cell segmentation that incorporates geometric features to improve performance on in situ tissue images using in vitro data.

Improving Generalizability of Protein Sequence Models with Data Augmentations

The paper improves protein sequence model generalizability by applying data augmentations and contrastive learning, showing performance gains across multiple tasks.

Harmonised segmentation of neonatal brain MRI

This paper applies unsupervised domain adaptation techniques to segment neonatal brain MRI data from a preterm population, improving segmentation accuracy and enabling analysis of cortical thickness and language outcomes.

Synthesizing VERDICT maps from standard DWI data using GANs

This study synthesizes VERDICT maps from standard DWI data using GANs, showing strong correlation between synthetic and real maps for tumor area quantification.

ACE: Explaining cluster from an adversarial perspective

ACE integrates clustering and marker gene selection into a single deep learning workflow to identify discriminative and nonredundant gene panels.