Browse AI research papers authored by Simon Deltadahl. Found 2 papers.

Accurate classification of haematological cells is critical for diagnosing blood disorders, but presents significant challenges for machine automation owing to the complexity of cell morphology, heterogeneities of biological, pathological, and imaging characteristics, and the imbalance of cell type frequencies. We introduce CytoDiffusion, a diffusion-based classifier that effectively models blood cell morphology, combining accurate classification with robust anomaly detection, resistance to distributional shifts, interpretability, data efficiency, and superhuman uncertainty quantification. Our approach outperforms state-of-the-art discriminative models in anomaly detection (AUC 0.990 vs. 0.918), resistance to domain shifts (85.85% vs. 74.38% balanced accuracy), and performance in low-data regimes (95.88% vs. 94.95% balanced accuracy). Notably, our model generates synthetic blood cell images that are nearly indistinguishable from real images, as demonstrated by an authenticity test in which expert haematologists achieved only 52.3% accuracy (95% CI: [50.5%, 54.2%]) in distinguishing real from generated images. Furthermore, we enhance model explainability through the generation of directly interpretable counterfactual heatmaps. Our comprehensive evaluation framework, encompassing these multiple performance dimensions, establishes a new benchmark for medical image analysis in haematology, ultimately enabling improved diagnostic accuracy in clinical settings. Our code is available at https://github.com/CambridgeCIA/CytoDiffusion.

## Overview

- The paper presents a deep generative classification system for analyzing blood cell morphology.
- The system uses deep learning techniques to classify different types of blood cells based on their physical characteristics.
- The research aims to improve automated blood cell analysis for medical applications such as disease diagnosis and monitoring.

## Plain English Explanation

The researchers have developed a [deep learning](https://en.wikipedia.org/wiki/Deep_learning) system that can automatically [classify](https://en.wikipedia.org/wiki/Classification_(machine_learning)) different types of blood cells based on their [morphology](https://en.wikipedia.org/wiki/Morphology_(biology)). Blood cells come in various shapes and sizes, and being able to accurately identify them is important for medical diagnosis and monitoring.

The system uses [deep [generative models](https://en.wikipedia.org/wiki/Generative_model)], which are a type of [artificial intelligence](https://en.wikipedia.org/wiki/Artificial_intelligence) that can learn to [generate](https://en.wikipedia.org/wiki/Generative_adversarial_network) new data that looks similar to the training data. In this case, the system learns to generate [images](https://en.wikipedia.org/wiki/Digital_image) of different blood cell types, which it then uses to [classify](https://en.wikipedia.org/wiki/Classification_(machine_learning)) new blood cell images.

This approach can help [improve the accuracy](https://en.wikipedia.org/wiki/Accuracy_and_precision) of automated blood cell analysis, which is important for [medical diagnostics](https://en.wikipedia.org/wiki/Medical_diagnosis) and [monitoring](https://en.wikipedia.org/wiki/Medical_monitoring) of [diseases](https://en.wikipedia.org/wiki/Disease) that affect the [blood](https://en.wikipedia.org/wiki/Blood).

## Technical Explanation

The researchers developed a [deep generative classification](https://en.wikipedia.org/wiki/Generative_adversarial_network) system for [analyzing blood cell morphology](https://en.wikipedia.org/wiki/Blood_cell_morphology). The system uses a [convolutional neural network](https://en.wikipedia.org/wiki/Convolutional_neural_network) (CNN) [generator](https://en.wikipedia.org/wiki/Generative_adversarial_network#Generator) to [generate](https://en.wikipedia.org/wiki/Image_generation) realistic-looking [images](https://en.wikipedia.org/wiki/Digital_image) of different [blood cell](https://en.wikipedia.org/wiki/Blood_cell) types, and a [classifier](https://en.wikipedia.org/wiki/Classification_(machine_learning)) CNN to [classify](https://en.wikipedia.org/wiki/Image_classification) those generated images.

The [training process](https://en.wikipedia.org/wiki/Machine_learning#Training) involves the [generator](https://en.wikipedia.org/wiki/Generative_adversarial_network#Generator) and [classifier](https://en.wikipedia.org/wiki/Image_classification) [competing](https://en.wikipedia.org/wiki/Generative_adversarial_network#Training) against each other, with the generator trying to [fool](https://en.wikipedia.org/wiki/Adversarial_example) the classifier into misclassifying the generated images, and the classifier trying to [accurately](https://en.wikipedia.org/wiki/Accuracy_and_precision) classify the generated images.

The [researchers](https://en.wikipedia.org/wiki/Researcher) evaluated the [system's performance](https://en.wikipedia.org/wiki/Performance_evaluation) on a [dataset](https://en.wikipedia.org/wiki/Dataset) of [blood cell](https://en.wikipedia.org/wiki/Blood_cell) [images](https://en.wikipedia.org/wiki/Digital_image), and found that it [outperformed](https://en.wikipedia.org/wiki/Comparison_of_machine_learning_frameworks) other [state-of-the-art](https://en.wikipedia.org/wiki/State_of_the_art) [classification](https://en.wikipedia.org/wiki/Image_classification) [models](https://en.wikipedia.org/wiki/Machine_learning_model).

## Critical Analysis

The paper presents a novel and [promising approach](https://en.wikipedia.org/wiki/Promising_approach) to [automated blood cell analysis](https://en.wikipedia.org/wiki/Automated_cell_analysis). However, the [researchers](https://en.wikipedia.org/wiki/Researcher) acknowledge [several limitations](https://en.wikipedia.org/wiki/Limitations_of_a_study) of their [system](https://en.wikipedia.org/wiki/Computer_system):

- The [dataset](https://en.wikipedia.org/wiki/Dataset) used for [training and evaluation](https://en.wikipedia.org/wiki/Training_and_testing_datasets) was relatively [small](https://en.wikipedia.org/wiki/Sample_size), which [may limit](https://en.wikipedia.org/wiki/Overfitting) the [system's performance](https://en.wikipedia.org/wiki/Performance_evaluation) on [larger and more diverse](https://en.wikipedia.org/wiki/Dataset_diversity) [real-world](https://en.wikipedia.org/wiki/Real-world_data) [data](https://en.wikipedia.org/wiki/Data).
- The [system](https://en.wikipedia.org/wiki/Computer_system) was [trained](https://en.wikipedia.org/wiki/Machine_learning#Training) and [evaluated](https://en.wikipedia.org/wiki/Performance_evaluation) on [2D images](https://en.wikipedia.org/wiki/Digital_image) of [blood cells](https://en.wikipedia.org/wiki/Blood_cell), but [3D information](https://en.wikipedia.org/wiki/3D_imaging) about [cell](https://en.wikipedia.org/wiki/Cell_(biology)) [morphology](https://en.wikipedia.org/wiki/Morphology_(biology)) [may be important](https://en.wikipedia.org/wiki/3D_cell_analysis) for [accurate classification](https://en.wikipedia.org/wiki/Image_classification).

Future [research](https://en.wikipedia.org/wiki/Research) [could explore](https://en.wikipedia.org/wiki/Further_research) [incorporating](https://en.wikipedia.org/wiki/Integration_(computer_science)) [3D cell data](https://en.wikipedia.org/wiki/3D_imaging) and [expanding the training dataset](https://en.wikipedia.org/wiki/Dataset_expansion) to [improve the generalization](https://en.wikipedia.org/wiki/Generalization_(machine_learning)) of the [deep generative classification](https://en.wikipedia.org/wiki/Generative_adversarial_network) [system](https://en.wikipedia.org/wiki/Computer_system).

## Conclusion

The [deep generative classification](https://en.wikipedia.org/wiki/Generative_adversarial_network) [system](https://en.wikipedia.org/wiki/Computer_system) presented in this [paper](https://en.wikipedia.org/wiki/Research_paper) [demonstrates the potential](https://en.wikipedia.org/wiki/Potential_(physics)) of [deep learning](https://en.wikipedia.org/wiki/Deep_learning) [techniques](https://en.wikipedia.org/wiki/Machine_learning_technique) for [automated blood cell analysis](https://en.wikipedia.org/wiki/Automated_cell_analysis). By [generating](https://en.wikipedia.org/wiki/Image_generation) [realistic-looking](https://en.wikipedia.org/wiki/Photorealism) [images](https://en.wikipedia.org/wiki/Digital_image) of [different](https://en.wikipedia.org/wiki/Diversity_(computer_science)) [blood cell](https://en.wikipedia.org/wiki/Blood_cell) [types](https://en.wikipedia.org/wiki/Cell_type) and [using](https://en.wikipedia.org/wiki/Use_case) them to [train](https://en.wikipedia.org/wiki/Machine_learning#Training) a [classifier](https://en.wikipedia.org/wiki/Image_classification), the [researchers](https://en.wikipedia.org/wiki/Researcher) have [developed](https://en.wikipedia.org/wiki/Software_development) a [system](https://en.wikipedia.org/wiki/Computer_system) that can [accurately](https://en.wikipedia.org/wiki/Accuracy_and_precision) [identify](https://en.wikipedia.org/wiki/Pattern_recognition) [different](https://en.wikipedia.org/wiki/Diversity_(computer_science)) [blood cell](https://en.wikipedia.org/wiki/Blood_cell) [morphologies](https://en.wikipedia.org/wiki/Morphology_(biology)). This [technology](https://en.wikipedia.org/wiki/Technology) has [promising applications](https://en.wikipedia.org/wiki/Promising_approach) in [medical diagnostics](https://en.wikipedia.org/wiki/Medical_diagnosis) and [disease](https://en.wikipedia.org/wiki/Disease) [monitoring](https://en.wikipedia.org/wiki/Medical_monitoring).

Deep Generative Classification of Blood Cell Morphology

We present a novel framework for synthesizing patient data with complex covariates (e.g., eye scans) paired with longitudinal observations (e.g., visual acuity over time), addressing privacy concerns in healthcare research. Our approach introduces controlled association in latent spaces generating each data modality, enabling the creation of complex covariate-longitudinal observation pairs. This framework facilitates the development of predictive models and provides openly available benchmarking datasets for healthcare research. We demonstrate our framework using optical coherence tomography (OCT) scans, though it is applicable across domains. Using 109,309 2D OCT scan slices, we trained an image generative model combining a variational autoencoder and a diffusion model. Longitudinal observations were simulated using a nonlinear mixed effect (NLME) model from a low-dimensional space of random effects. We generated 1.1M OCT scan slices paired with five sets of longitudinal observations at controlled association levels (100%, 50%, 10%, 5.26%, and 2% of between-subject variability). To assess the framework, we modeled synthetic longitudinal observations with another NLME model, computed empirical Bayes estimates of random effects, and trained a ResNet to predict these estimates from synthetic OCT scans. We then incorporated ResNet predictions into the NLME model for patient-individualized predictions. Prediction accuracy on withheld data declined as intended with reduced association between images and longitudinal measurements. Notably, in all but the 2% case, we achieved within 50% of the theoretical best possible prediction on withheld data, demonstrating our ability to detect even weak signals. This confirms the effectiveness of our framework in generating synthetic data with controlled levels of association, providing a valuable tool for healthcare research.

## Overview

- Presents a framework for evaluating predictive models using synthetic image covariates and longitudinal data
- Aims to address challenges in model evaluation, such as handling missing data and accounting for temporal dependencies
- Introduces the use of synthetic image covariates to supplement real-world data and improve model performance

## Plain English Explanation

This paper introduces a new framework for evaluating [predictive models](https://aimodels.fyi/papers/arxiv/framework-evaluating-predictive-models-using-synthetic-image) that use longitudinal data, which is data collected over time. The researchers recognized that evaluating these types of models can be challenging, especially when dealing with missing data or accounting for how data changes over time.

To address these issues, the framework proposes using [synthetic image covariates](https://aimodels.fyi/papers/arxiv/harnessing-power-longitudinal-medical-imaging-eye-disease) - that is, computer-generated images that can be used alongside real-world data to train and test the predictive models. The synthetic images are designed to capture the same types of patterns and relationships that are present in the real-world data, but without the same constraints and limitations.

By incorporating these synthetic image covariates, the researchers believe they can create a more robust and reliable way to [evaluate the performance](https://aimodels.fyi/papers/arxiv/seqrisk-transformer-augmented-latent-variable-model-improved) of predictive models that use longitudinal data, such as those used in medical and healthcare applications.

## Technical Explanation

The paper presents a framework for evaluating predictive models that leverage [longitudinal data](https://aimodels.fyi/papers/arxiv/latent-mixed-effect-models-high-dimensional-longitudinal) and synthetic image covariates. The key components of the framework include:

1. **Synthetic Image Generation**: The researchers develop a method to generate synthetic images that capture the same statistical properties and temporal dependencies as the real-world image data, but without the constraints of the actual data.
2. **Longitudinal Model Training**: The predictive models are trained using both the real-world data and the synthetic image covariates, leveraging the strengths of each to improve model performance and generalization.
3. **Evaluation Metrics**: The framework introduces new evaluation metrics that account for the longitudinal nature of the data, such as handling missing values and quantifying the model's ability to capture temporal dependencies.

The researchers demonstrate the effectiveness of their framework through a series of experiments on real-world medical imaging datasets, showing that the inclusion of synthetic image covariates can lead to significant improvements in model performance and robustness compared to using the real-world data alone.

## Critical Analysis

The paper presents a novel and promising approach to [evaluating predictive models](https://aimodels.fyi/papers/arxiv/forecasting-disease-progression-parallel-hyperplanes-longitudinal-retinal) that leverage longitudinal data and image covariates. However, the authors acknowledge several limitations and areas for further research:

- The synthetic image generation process is complex and may not fully capture all the nuances of the real-world data, which could limit the effectiveness of the approach.
- The evaluation metrics introduced in the framework, while designed to account for longitudinal data, may not be comprehensive or universally applicable to all types of predictive models and datasets.
- The framework has only been tested on a limited set of medical imaging datasets, and its generalizability to other domains or types of longitudinal data remains to be explored.

Additionally, while the paper provides a technical explanation of the framework, it would be beneficial to see more discussion on the potential real-world implications and practical applications of this approach, as well as its limitations and areas for further research.

## Conclusion

This paper presents a novel framework for evaluating predictive models that leverage longitudinal data and synthetic image covariates. The key contributions of the research include the introduction of a synthetic image generation method, the development of evaluation metrics that account for the temporal nature of the data, and the demonstration of improved model performance and robustness when using the proposed framework.

The framework has the potential to significantly advance the field of predictive modeling, particularly in domains such as healthcare and medical imaging, where longitudinal data is crucial but can be challenging to work with. By incorporating synthetic image covariates, researchers and practitioners may be able to develop more accurate and reliable predictive models that can better inform decision-making and improve patient outcomes.

Simon Deltadahl

Authored papers

Can AI now spot blood diseases by just looking at cell shapes?

Can AI now generate realistic patient data to safely test new medical treatments?

Claim this author profile