5 May 2023 in Medical Image Analysis
Image registration aims to find geometric transformations that align images. Most algorithmic and deep learning-based methods solve the registration problem by minimizing a loss function, consisting of a similarity metric comparing the aligned images, and a regularization term ensuring smoothness of the transformation. Existing similarity metrics like Euclidean Distance or Normalized Cross-Correlation focus on aligning pixel intensity values or correlations, giving difficulties with low intensity contrast, noise, and ambiguous matching. We propose a semantic similarity metric for image registration, focusing on aligning image areas based on semantic correspondence instead. Our approach learns dataset-specific features that drive the optimization of a learning-based registration model. We train both an unsupervised approach extracting features with an auto-encoder, and a semi-supervised approach using supplemental segmentation data. We validate the semantic similarity metric using both deep-learning-based and algorithmic image registration methods. Compared to existing methods across four different image modalities and applications, the method achieves consistently high registration accuracy and smooth transformation fields.
30 November 2025 in medRxiv
Study objectives: To evaluate an AI system for musculoskeletal (MSK) radiography in confidently identifying examinations without injury-related pathologies to support AI-guided discharge of patients without traumatic findings. Methods: We retrospectively sampled radiographic examinations, including one or more radiographs and a radiological report, of suspected MSK trauma from > 1,000 clinical sites in two countries. Medically trained professionals independently classified all exams. When disagreements between the classification and the original radiological report arose, adjudication was performed by a third professional. Annotators were also asked to record their confidence level of each classification. The AI system analyzed all exams and assigned them to the categories AI Positive, AI Negative , and AI Negative (very high confidence) . Performance of the AI system was assessed using error rate, false negative rate (FNR), and qualitative review of misclassified exams. Results: A total of 2,962 exams were included. The AI classified 27.6% (818/2,961; 95% CI: 26.0–29.3) of exams as AI Negative (very high confidence) . Of all exams, 0.7% (21/2,962; 95% CI: 0.0–1.0) were falsely classified as highconfidence negatives, corresponding to a false negative rate (FNR) of 2.0% (21/1,026; 95% CI: 0.0–2.9). Qualitative review of false negatives showed that the majority had no clinical consequence if correctly diagnosed during routine follow-up the following day, and no clearly high-risk exams were missed. Conclusion: The AI system identified over one-quarter of MSK trauma radiographs as confidently negative with a very low rate of false negatives, performing on par or better than the reported standard-of-care. These results suggest potential for safe, AI-driven decision support and workflow optimization for the discharge of patients with clearly negative examinations.