Draft Status
Ready - team will start page creating immediately
Category
Segmentation / Classification / Landmarking
Key Investigators
- Deepa Krishnaswamy (Brigham and Women's Hospital, USA)
- Alexandre Banks Gadbois (Brigham and Women's Hospital, USA)
- Dave Dinh (Consultant for Brigham and Women's Hospital, USA)
- Matt McCormick (Fideus Labs, USA)
- Tina Kapur (Brigham and Women's Hospital, USA)
Project Description
One of the primary indicators of acute heart failure is the presence of pulmonary congestion. To detect the fluid build up quickly the in emergency room, lung ultrasound exams are taken of the patient. Clinicians look for hyperechoic artifacts (B-lines) that appear in the image, where the more they see, the more congested the patient is.
Unfortunately, the manual detection of these B-lines is difficult due to the image quality, which depends on the type of transducer and the expertise of the clinician. Therefore, AI models have started to be developed to quickly detect these B-lines. Our group, over the past few years, has developed multiple methods [1], [2], [3]. One of the drawbacks though, is that the models do not compute the pleural percentage, which is the ratio of the B-line sectors to the pleural line sectors.
Therefore, we have started to investigate using object detection models for automatically finding the pleural lines and B-lines. Our goal for this project week is to train and validate detection models such as YOLO, [4] and talk to clinicians about the performance.
Objective
- Train and validate YOLO models for pleural line and B-line detection.
- Train and validate other state-of-the-art methods, such as RF-DETR [5]
- Show and discuss our results with clinicians.
Approach and Plan
- Develop models for pleural line and B-line detection.
- Make our model/code publicly available.
- Talk to clinicians
Progress and Next Steps
- We have trained some preliminary YOLO models. We trained one oriented bounding box YOLO model for pleural lines, and another axis-aligned box YOLO model for B-lines.
- We calculated the percent pleura and compared it to the experts.
Illustrations
Ground truth boxes in scanline space: oriented bounding box for the pleural line (left), and axis-aligned bounding box for the B-lines.
YOLO results after performing pleural line detection and B-line detection, and conversion back to the sector space, compared to the ground truth:
Percent pleura for one test lung zone - x = frame #, and y = percentage. We see that the AI predicted
Background and References
[1] Lucassen RT, Jafari MH, Duggan NM, Jowkar N, Mehrtash A, Fischetti C, Bernier D, Prentice K, Duhaime EP, Jin M, Abolmaesumi P. Deep learning for detection and localization of B-lines in lung ultrasound. IEEE journal of biomedical and health informatics. 2023 Jun 5;27(9):4352-61.
[2] Asgari-Targhi A, Ungi T, Jin M, Harrison N, Duggan N, Duhaime E, Goldsmith A, Kapur T. Can Crowdsourced Annotations Improve AI-Based Congestion Scoring for Bedside Lung Ultrasound?. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention 2024 Oct 7 (pp. 580-590). Cham: Springer Nature Switzerland.[3] MICCAI 2026 submission.
[3] MICCAI 2026 acceptance. AI-Driven Pulmonary Congestion Assessment for Lung Ultrasound via Segmentation-Guided Transformers.
[4] https://github.com/ultralytics/ultralytics
[5] https://github.com/roboflow/rf-detr
Draft Status
Ready - team will start page creating immediately
Category
Segmentation / Classification / Landmarking
Key Investigators
Project Description
One of the primary indicators of acute heart failure is the presence of pulmonary congestion. To detect the fluid build up quickly the in emergency room, lung ultrasound exams are taken of the patient. Clinicians look for hyperechoic artifacts (B-lines) that appear in the image, where the more they see, the more congested the patient is.
Unfortunately, the manual detection of these B-lines is difficult due to the image quality, which depends on the type of transducer and the expertise of the clinician. Therefore, AI models have started to be developed to quickly detect these B-lines. Our group, over the past few years, has developed multiple methods [1], [2], [3]. One of the drawbacks though, is that the models do not compute the pleural percentage, which is the ratio of the B-line sectors to the pleural line sectors.
Therefore, we have started to investigate using object detection models for automatically finding the pleural lines and B-lines. Our goal for this project week is to train and validate detection models such as YOLO, [4] and talk to clinicians about the performance.
Objective
Approach and Plan
Progress and Next Steps
Illustrations
Ground truth boxes in scanline space: oriented bounding box for the pleural line (left), and axis-aligned bounding box for the B-lines.
YOLO results after performing pleural line detection and B-line detection, and conversion back to the sector space, compared to the ground truth:
Percent pleura for one test lung zone - x = frame #, and y = percentage. We see that the AI predicted
Background and References
[1] Lucassen RT, Jafari MH, Duggan NM, Jowkar N, Mehrtash A, Fischetti C, Bernier D, Prentice K, Duhaime EP, Jin M, Abolmaesumi P. Deep learning for detection and localization of B-lines in lung ultrasound. IEEE journal of biomedical and health informatics. 2023 Jun 5;27(9):4352-61.
[2] Asgari-Targhi A, Ungi T, Jin M, Harrison N, Duggan N, Duhaime E, Goldsmith A, Kapur T. Can Crowdsourced Annotations Improve AI-Based Congestion Scoring for Bedside Lung Ultrasound?. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention 2024 Oct 7 (pp. 580-590). Cham: Springer Nature Switzerland.[3] MICCAI 2026 submission.
[3] MICCAI 2026 acceptance. AI-Driven Pulmonary Congestion Assessment for Lung Ultrasound via Segmentation-Guided Transformers.
[4] https://github.com/ultralytics/ultralytics
[5] https://github.com/roboflow/rf-detr