Deep Learning Based Emboli Detection Using Ultrasound Doppler Imaging

Raghava Vinaykanth Mushunuri, Cecilie le Duc Dahl, Elisabeth Krogstad Iversen, Sigrid Dannheim Vik, Martin Letho-Olsen, Hans Torp, Siri Ann Nyrnes, Gabriel Kiss

Keywords: Emboli detection, NeoDoppler, Cardio Vascular Intervention, Explainable AI, High Intensity Transient Signals, Deep Learning, Transfer Learning

TL;DR: This manuscript deals with detection of emboli signals reaching to the brain during cardiovascular interventions/ surgeries using state-of-the-art deep learning models. We evaluated the performance using unseen test cases and Explainable AI methods.

Abstract: Accurate detection of embolic signals in the bloodstream is crucial for early diagnosis and prevention of cerebrovascular complications, and this work develops and evaluates an artificial intelligence–based system for automatic emboli detection in power Doppler imaging from NeoDoppler, aiming for robust and real-time performance. The study uses a four-stage experimental pipeline built on convolutional neural networks with transfer learning: an initial baseline model (Stage 1), an assessment of spatial generalisation (Stage 2), and a hybrid two-step strategy (Stage 3) that combines conventional High-Intensity Transient Signal (HITS) pre-detection with CNN-based classification, followed by a simplified preprocessing strategy in Stage 4, where single-channel images are replicated into three channels to match pre-trained CNN architectures; all models are trained with 5-fold cross-validation on 523 recordings from 25 patients and evaluated on unseen pilot recordings from the same cohort and additional abdominal surgery data. Across stages, performance improves progressively, with the hybrid two-step framework using the three-channel replication yielding strong results, achieving 96% sensitivity and 98% specificity on the pilot recording and 94% sensitivity and 71% specificity on the abdominal surgery recordings. We estimated 95% confidence intervals (CIs) using Wilson's score for abdominal surgery recordings, with a CI of 0.730-0.99, demonstrating that the proposed approach is an efficient and interpretable solution for ultrasound-based emboli monitoring.

Primary Subject Area: Application: Cardiology

Secondary Subject Area: Transfer Learning and Domain Adaptation

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 19