Evaluating Reliability in Medical DNNs: A Critical Analysis of Feature and Confidence-Based OOD Detection

Reliable use of deep neural networks (DNNs) for medical image analysis requires methods to identify inputs that differ significantly from the training data, called out-of-distribution (OOD), to prevent erroneous predictions. OOD detection methods can be categorised as either confidence-based (using the model's output layer for OOD detection) or feature-based (not using the output layer). We created two new OOD benchmarks by dividing the D7P (dermatology) and BreastMNIST (ultrasound) datasets into subsets which either contain or don't contain an artefact (rulers or annotations respectively). Models were trained with artefact-free images, and images with the artefacts were used as OOD test sets. For each OOD image, we created a counterfactual by manually removing the artefact via image processing, to assess the artefact's impact on the model's predictions. We show that OOD artefacts can boost a model's softmax confidence in its predictions, due to correlations in training data among other factors. This contradicts the common assumption that OOD artefacts should lead to more uncertain outputs, an assumption on which most confidence-based methods rely. We use this to explain why feature-based methods (e.g. Mahalanobis score) typically have greater OOD detection performance than confidence-based methods (e.g. MCP). However, we also show that feature-based methods typically perform worse at distinguishing between inputs that lead to correct and incorrect predictions (for both OOD and ID data). Following from these insights, we argue that a combination of feature-based and confidence-based methods should be used within DNN pipelines to mitigate their respective weaknesses. These project's code and OOD benchmarks are available at: https://github.com/HarryAnthony/Evaluating_OOD_detection.

Quality Control for Radiology Report Generation Models via Auxiliary Auditing Components

Automation of medical image interpretation could alleviate bottlenecks in diagnostic workflows, and has become of particular interest in recent years due to advancements in natural language processing. Great strides have been made towards automated radiology report generation via AI, yet ensuring clinical accuracy in generated reports is a significant challenge, hindering deployment of such methods in clinical practice. In this work we propose a quality control framework for assessing the reliability of AI-generated radiology reports with respect to semantics of diagnostic importance using modular auxiliary auditing components (AC). Evaluating our pipeline on the MIMIC-CXR dataset, our findings show that incorporating ACs in the form of disease-classifiers can enable auditing that identifies more reliable reports, resulting in higher F1 scores compared to unfiltered generated reports. Additionally, leveraging the confidence of the AC labels further improves the audit's effectiveness.

An Organism Starts with a Single Pix-Cell: A Neural Cellular Diffusion for High-Resolution Image Synthesis

Generative modeling seeks to approximate the statistical properties of real data, enabling synthesis of new data that closely resembles the original distribution. Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs) represent significant advancements in generative modeling, drawing inspiration from game theory and thermodynamics, respectively. Nevertheless, the exploration of generative modeling through the lens of biological evolution remains largely untapped. In this paper, we introduce a novel family of models termed Generative Cellular Automata (GeCA), inspired by the evolution of an organism from a single cell. GeCAs are evaluated as an effective augmentation tool for retinal disease classification across two imaging modalities: Fundus and Optical Coherence Tomography (OCT). In the context of OCT imaging, where data is scarce and the distribution of classes is inherently skewed, GeCA significantly boosts the performance of 11 different ophthalmological conditions, achieving a 12% increase in the average F1 score compared to conventional baselines. GeCAs outperform both diffusion methods that incorporate UNet or state-of-the art variants with transformer-based denoising models, under similar parameter constraints. Code is available at: https://github.com/xmed-lab/GeCA.

Feasibility of Federated Learning from Client Databases with Different Brain Diseases and MRI Modalities

Segmentation models for brain lesions in MRI are commonly developed for a specific disease and trained on data with a predefined set of MRI modalities. Each such model cannot segment the disease using data with a different set of MRI modalities, nor can it segment any other type of disease. Moreover, this training paradigm does not allow a model to benefit from learning from heterogeneous databases that may contain scans and segmentation labels for different types of brain pathologies and diverse sets of MRI modalities. Is it feasible to use Federated Learning (FL) for training a single model on client databases that contain scans and labels of different brain pathologies and diverse sets of MRI modalities? We demonstrate promising results by combining appropriate, simple, and practical modifications to the model and training strategy: Designing a model with input channels that cover the whole set of modalities available across clients, training with random modality drop, and exploring the effects of feature normalization methods. Evaluation on 7 brain MRI databases with 5 different diseases shows that such FL framework can train a single model that is shown to be very promising in segmenting all disease types seen during training. Importantly, it is able to segment these diseases in new databases that contain sets of modalities different from those in training clients. These results demonstrate, for the first time, feasibility and effectiveness of using FL to train a single segmentation model on decentralised data with diverse brain diseases and MRI modalities, a necessary step towards leveraging heterogeneous real-world databases. Code will be made available at: https://github.com/FelixWag/FL-MultiDisease-MRI

IterMask2: Iterative Unsupervised Anomaly Segmentation via Spatial and Frequency Masking for Brain Lesions in MRI

Unsupervised anomaly segmentation approaches to pathology segmentation train a model on images of healthy subjects, that they define as the 'normal' data distribution. At inference, they aim to segment any pathologies in new images as 'anomalies', as they exhibit patterns that deviate from those in 'normal' training data. Prevailing methods follow the 'corrupt-and-reconstruct' paradigm. They intentionally corrupt an input image, reconstruct it to follow the learned 'normal' distribution, and subsequently segment anomalies based on reconstruction error. Corrupting an input image, however, inevitably leads to suboptimal reconstruction even of normal regions, causing false positives. To alleviate this, we propose a novel iterative spatial mask-refining strategy IterMask2. We iteratively mask areas of the image, reconstruct them, and update the mask based on reconstruction error. This iterative process progressively adds information about areas that are confidently normal as per the model. The increasing content guides reconstruction of nearby masked areas, improving reconstruction of normal tissue under these areas, reducing false positives. We also use high-frequency image content as an auxiliary input to provide additional structural information for masked areas. This further improves reconstruction error of normal in comparison to anomalous areas, facilitating segmentation of the latter. We conduct experiments on several brain lesion datasets and demonstrate effectiveness of our method. Code is available at: https://github.com/ZiyunLiang/IterMasks2

Feasibility and benefits of joint learning from MRI databases with different brain diseases and modalities for segmentation

Models for segmentation of brain lesions in multi-modal MRI are commonly trained for a specific pathology using a single database with a predefined set of MRI modalities, determined by a protocol for the specific disease. This work explores the following open questions: Is it feasible to train a model using multiple databases that contain varying sets of MRI modalities and annotations for different brain pathologies? Will this joint learning benefit performance on the sets of modalities and pathologies available during training? Will it enable analysis of new databases with different sets of modalities and pathologies? We develop and compare different methods and show that promising results can be achieved with appropriate, simple and practical alterations to the model and training framework. We experiment with 7 databases containing 5 types of brain pathologies and different sets of MRI modalities. Results demonstrate, for the first time, that joint training on multi-modal MRI databases with different brain pathologies and sets of modalities is feasible and offers practical benefits. It enables a single model to segment pathologies encountered during training in diverse sets of modalities, while facilitating segmentation of new types of pathologies such as via follow-up fine-tuning. The insights this study provides into the potential and limitations of this paradigm should prove useful for guiding future advances in the direction. Code and pretrained models: https://github.com/WenTXuL/MultiUnet

Semi-Supervised Learning for Deep Causal Generative Models

Developing models that can answer questions of the form "How would $x$ change if $y$ had been $z$?" is fundamental for advancing medical image analysis. Training causal generative models that address such counterfactual questions, though, currently requires that all relevant variables have been observed and that corresponding labels are available in training data. However, clinical data may not have complete records for all patients and state of the art causal generative models are unable to take full advantage of this. We thus develop, for the first time, a semi-supervised deep causal generative model that exploits the causal relationships between variables to maximise the use of all available data. We explore this in the setting where each sample is either fully labelled or fully unlabelled, as well as the more clinically realistic case of having different labels missing for each sample. We leverage techniques from causal inference to infer missing values and subsequently generate realistic counterfactuals, even for samples with incomplete labels.

As Firm As Their Foundations: Can open-sourced foundation models be used to create adversarial examples for downstream tasks?

Foundation models pre-trained on web-scale vision-language data, such as CLIP, are widely used as cornerstones of powerful machine learning systems. While pre-training offers clear advantages for downstream learning, it also endows downstream models with shared adversarial vulnerabilities that can be easily identified through the open-sourced foundation model. In this work, we expose such vulnerabilities in CLIP's downstream models and show that foundation models can serve as a basis for attacking their downstream systems. In particular, we propose a simple yet effective adversarial attack strategy termed Patch Representation Misalignment (PRM). Solely based on open-sourced CLIP vision encoders, this method produces adversaries that simultaneously fool more than 20 downstream models spanning 4 common vision-language tasks (semantic segmentation, object detection, image captioning and visual question-answering). Our findings highlight the concerning safety risks introduced by the extensive usage of public foundational models in the development of downstream systems, calling for extra caution in these scenarios.

Examining Modality Incongruity in Multimodal Federated Learning for Medical Vision and Language-based Disease Detection

Multimodal Federated Learning (MMFL) utilizes multiple modalities in each client to build a more powerful Federated Learning (FL) model than its unimodal counterpart. However, the impact of missing modality in different clients, also called modality incongruity, has been greatly overlooked. This paper, for the first time, analyses the impact of modality incongruity and reveals its connection with data heterogeneity across participating clients. We particularly inspect whether incongruent MMFL with unimodal and multimodal clients is more beneficial than unimodal FL. Furthermore, we examine three potential routes of addressing this issue. Firstly, we study the effectiveness of various self-attention mechanisms towards incongruity-agnostic information fusion in MMFL. Secondly, we introduce a modality imputation network (MIN) pre-trained in a multimodal client for modality translation in unimodal clients and investigate its potential towards mitigating the missing modality problem. Thirdly, we assess the capability of client-level and server-level regularization techniques towards mitigating modality incongruity effects. Experiments are conducted under several MMFL settings on two publicly available real-world datasets, MIMIC-CXR and Open-I, with Chest X-Ray and radiology reports.

On the use of Mahalanobis distance for out-of-distribution detection with neural networks for medical imaging

Implementing neural networks for clinical use in medical applications necessitates the ability for the network to detect when input data differs significantly from the training data, with the aim of preventing unreliable predictions. The community has developed several methods for out-of-distribution (OOD) detection, within which distance-based approaches - such as Mahalanobis distance - have shown potential. This paper challenges the prevailing community understanding that there is an optimal layer, or combination of layers, of a neural network for applying Mahalanobis distance for detection of any OOD pattern. Using synthetic artefacts to emulate OOD patterns, this paper shows the optimum layer to apply Mahalanobis distance changes with the type of OOD pattern, showing there is no one-fits-all solution. This paper also shows that separating this OOD detector into multiple detectors at different depths of the network can enhance the robustness for detecting different OOD patterns. These insights were validated on real-world OOD tasks, training models on CheXpert chest X-rays with no support devices, then using scans with unseen pacemakers (we manually labelled 50% of CheXpert for this research) and unseen sex as OOD cases. The results inform best-practices for the use of Mahalanobis distance for OOD detection. The manually annotated pacemaker labels and the project's code are available at: https://github.com/HarryAnthony/Mahalanobis-OOD-detection.