19 September 2022 in IEEE Journal of Biomedical and Health Informatics
Recently a number of studies demonstrated impressive performance on diverse vision-language multi-modal tasks such as image captioning and visual question answering by extending the BERT architecture with multi-modal pre-training objectives. In this work we explore a broad set of multi-modal representation learning tasks in the medical domain, specifically using radiology images and the unstructured report. We propose Medical Vision Language Learner (MedViLL), which adopts a BERT-based architecture combined with a novel multi-modal attention masking scheme to maximize generalization performance for both vision-language understanding tasks (diagnosis classification, medical image-report retrieval, medical visual question answering) and vision-language generation task (radiology report generation). By statistically and rigorously evaluating the proposed model on four downstream tasks with three radiographic image-report datasets (MIMIC-CXR, Open-I, and VQA-RAD), we empirically demonstrate the superior downstream task performance of MedViLL against various baselines, including task-specific architectures. The source code is publicly available at: https://github.com/SuperSupermoon/MedViLL
1 January 2021 in Computer Methods and Programs in Biomedicine
Coronary artery disease, which is mostly caused by atherosclerotic narrowing of the coronary artery lumen, is a leading cause of death. Coronary angiography is the standard method to estimate the severity of coronary artery stenosis, but is frequently limited by intra- and inter-observer variations. We propose a deep-learning algorithm that automatically recognizes stenosis in coronary angiographic images. The proposed method consists of key frame detection, deep learning model training for classification of stenosis on each key frame, and visualization of the possible location of the stenosis. Firstly, we propose an algorithm that automatically extracts key frames essential for diagnosis from 452 right coronary artery angiography movie clips. Our deep learning model is then trained with image-level annotations to classify the areas narrowed by over 50 %. To make the model focus on the salient features, we apply a self-attention mechanism. The stenotic locations are visualized using the activated area of feature maps with gradient-weighted class activation mapping. The automatically detected key frame was very close to the manually selected key frame (average distance (1.70 ± 0.12) frame per clip). The model was trained with key frames on internal datasets, and validated with internal and external datasets. Our training method achieved high frame-wise area-under-the-curve of 0.971, frame-wise accuracy of 0.934, and clip-wise accuracy of 0.965 in the average values of cross-validation evaluations. The external validation results showed high performances with the mean frame-wise area-under-the-curve of (0.925 and 0.956) in the single and ensemble model, respectively. Heat map visualization shows the location for different types of stenosis in both internal and external data sets. With the self-attention mechanism, the stenosis could be precisely localized, which helps to accurately classify the stenosis by type. Our automated classification algorithm could recognize and localize coronary artery stenosis highly accurately. Our approach might provide the basis for a screening and assistant tool for the interpretation of coronary angiography.
29 June 2021 in JMIR Preprints
BACKGROUND Recently a number of studies demonstrated impressive performance on diverse vision-language multi-modal tasks such as image captioning and visual question answering by extending the self-attention based Transformer architecture with multi-modal pre-training objectives. Despite its huge potential, vision-language multi-modal pre-training in the medical domain has only recently received attention, only demonstrated improved diagnosis accuracy of vision-language pre-trained models. OBJECTIVE In this work we explore a broad set of multi-modal representation learning tasks in the medical domain, specifically using radiology images and the unstructured report. METHODS We propose Medical Vision Language Learner (MedViLL) which adopts a Transformer-based architecture combined with a novel multi-modal attention masking scheme to maximize generalization performance for both vision-language understanding tasks (disease classification, medical image-report retrieval, and medical visual question answering) and vision-language generation task (radiology report generation). RESULTS By rigorously evaluating the proposed model on four downstream tasks with three radiographic image-text datasets (MIMIC-CXR, Open-I, and VQA-RAD), we empirically demonstrate the superior downstream task performance and generality of MedViLL against various baselines including task-specific architectures. In addition, we qualitatively analyze MedViLL by showing the results of retrieved image-report pair, the attention map visualization, and generated reports. CONCLUSIONS Our proposed multi-modal pre-training model MedViLL could flexibly adapt to multiple downstream tasks of vision-language understanding and generation with a novel self-attention scheme. We believe that our approach can provide the basis for a wide range of interpretations of vision-language multi-modal in the medical domain.
1 January 2025 in Yonsei Medical Journal
Lee SJ, et al. Yonsei Med J. 2025 May;66:e45. https://doi.org/10.3349/ymj.2024.0246