Theresa Willem

Alessandro Wollek, Robert Graf, Saša Čečatka et al.

Purpose: To investigate chest radiograph (CXR) classification performance of vision transformers (ViT) and interpretability of attention-based saliency using the example of pneumothorax classification. Materials and Methods: In this retrospective study, ViTs were fine-tuned for lung disease classification using four public data sets: CheXpert, Chest X-Ray 14, MIMIC CXR, and VinBigData. Saliency maps were generated using transformer multimodal explainability and gradient-weighted class activation mapping (GradCAM). Classification performance was evaluated on the Chest X-Ray 14, VinBigData, and SIIM-ACR data sets using the area under the receiver operating characteristic curve analysis (AUC) and compared with convolutional neural networks (CNNs). The explainability methods were evaluated with positive/negative perturbation, sensitivity-n, effective heat ratio, intra-architecture repeatability and interarchitecture reproducibility. In the user study, three radiologists classified 160 CXRs with/without saliency maps for pneumothorax and rated their usefulness. Results: ViTs had comparable CXR classification AUCs compared with state-of-the-art CNNs 0.95 (95% CI: 0.943, 0.950) versus 0.83 (95%, CI 0.826, 0.842) on Chest X-Ray 14, 0.84 (95% CI: 0.769, 0.912) versus 0.83 (95% CI: 0.760, 0.895) on VinBigData, and 0.85 (95% CI: 0.847, 0.861) versus 0.87 (95% CI: 0.868, 0.882) on SIIM ACR. Both saliency map methods unveiled a strong bias toward pneumothorax tubes in the models. Radiologists found 47% of the attention-based saliency maps useful and 39% of GradCAM. The attention-based methods outperformed GradCAM on all metrics. Conclusion: ViTs performed similarly to CNNs in CXR classification, and their attention-based saliency maps were more useful to radiologists and outperformed GradCAM.

Alessandro Wollek, Theresa Willem, Michael Ingrisch et al.

To investigate the impact of OOD radiographs on existing chest X-ray classification models and to increase their robustness against OOD data. The study employed the commonly used chest X-ray classification model, CheXnet, trained on the chest X-ray 14 data set, and tested its robustness against OOD data using three public radiography data sets: IRMA, Bone Age, and MURA, and the ImageNet data set. To detect OOD data for multi-label classification, we proposed in-distribution voting (IDV). The OOD detection performance is measured across data sets using the area under the receiver operating characteristic curve (AUC) analysis and compared with Mahalanobis-based OOD detection, MaxLogit, MaxEnergy and self-supervised OOD detection (SS OOD). Without additional OOD detection, the chest X-ray classifier failed to discard any OOD images, with an AUC of 0.5. The proposed IDV approach trained on ID (chest X-ray 14) and OOD data (IRMA and ImageNet) achieved, on average, 0.999 OOD AUC across the three data sets, surpassing all other OOD detection methods. Mahalanobis-based OOD detection achieved an average OOD detection AUC of 0.982. IDV trained solely with a few thousand ImageNet images had an AUC 0.913, which was higher than MaxLogit (0.726), MaxEnergy (0.724), and SS OOD (0.476). The performance of all tested OOD detection methods did not translate well to radiography data sets, except Mahalanobis-based OOD detection and the proposed IDV method. Training solely on ID data led to incorrect classification of OOD images as ID, resulting in increased false positive rates. IDV substantially improved the model's ID classification performance, even when trained with data that will not occur in the intended use case or test set, without additional inference overhead.

Theodor Cheslerean-Boghiu, Melia-Evelina Fleischmann, Theresa Willem et al.

A lot of deep learning (DL) research these days is mainly focused on improving quantitative metrics regardless of other factors. In human-centered applications, like skin lesion classification in dermatology, DL-driven clinical decision support systems are still in their infancy due to the limited transparency of their decision-making process. Moreover, the lack of procedures that can explain the behavior of trained DL algorithms leads to almost no trust from clinical physicians. To diagnose skin lesions, dermatologists rely on visual assessment of the disease and the data gathered from the patient's anamnesis. Data-driven algorithms dealing with multi-modal data are limited by the separation of feature-level and decision-level fusion procedures required by convolutional architectures. To address this issue, we enable single-stage multi-modal data fusion via the attention mechanism of transformer-based architectures to aid in diagnosing skin diseases. Our method beats other state-of-the-art single- and multi-modal DL architectures in image-rich and patient-data-rich environments. Additionally, the choice of the architecture enables native interpretability support for the classification task both in the image and metadata domain with no additional modifications necessary.

Haroui Ma, Francesco Quinzan, Theresa Willem et al.

Machine learning (ML) systems for medical imaging have demonstrated remarkable diagnostic capabilities, but their susceptibility to biases poses significant risks, since biases may negatively impact generalization performance. In this paper, we introduce a novel statistical framework to evaluate the dependency of medical imaging ML models on sensitive attributes, such as demographics. Our method leverages the concept of counterfactual invariance, measuring the extent to which a model's predictions remain unchanged under hypothetical changes to sensitive attributes. We present a practical algorithm that combines conditional latent diffusion models with statistical hypothesis testing to identify and quantify such biases without requiring direct access to counterfactual data. Through experiments on synthetic datasets and large-scale real-world medical imaging datasets, including \textsc{cheXpert} and MIMIC-CXR, we demonstrate that our approach aligns closely with counterfactual fairness principles and outperforms standard baselines. This work provides a robust tool to ensure that ML diagnostic systems generalize well, e.g., across demographic groups, offering a critical step towards AI safety in healthcare. Code: https://github.com/Neferpitou3871/AI-Alignment-Medical-Imaging.