Multimedia Tools and Applications, Год журнала: 2024, Номер unknown
Опубликована: Сен. 16, 2024
Язык: Английский
Multimedia Tools and Applications, Год журнала: 2024, Номер unknown
Опубликована: Сен. 16, 2024
Язык: Английский
Knowledge-Based Systems, Год журнала: 2025, Номер unknown, С. 113538 - 113538
Опубликована: Апрель 1, 2025
Процитировано
0Journal of Biomedical Informatics, Год журнала: 2024, Номер 161, С. 104764 - 104764
Опубликована: Дек. 31, 2024
Язык: Английский
Процитировано
1ACM Transactions on Asian and Low-Resource Language Information Processing, Год журнала: 2024, Номер 23(9), С. 1 - 17
Опубликована: Июль 3, 2024
Increase in Chest X-ray (CXR) imaging tests has burdened radiologists, thereby posing significant challenges writing radiological reports on time. Although several deep learning-based automatic report generation methods have been developed, most are over-parameterized. For deployment edge devices with constrained processing power or limited resources, over-parameterized models often too large. This article presents a compressed model that is 30% space efficient compared to the non-compressed base model, while both comparable performance. The comprising VGG19 and hierarchical long short-term memory equipped contextual word embedding layer used as model. redundant weight parameters removed from using unstructured one-shot pruning. To overcome performance degradation, lightweight pruned fine-tuned over publicly available OpenI dataset. quantitative evaluation metric scores demonstrate proposed surpasses of state-of-the-art models. Additionally, being efficient, easily deployable resource-limited settings. Thus, this study serves baseline for development generate CXR images.
Язык: Английский
Процитировано
0Electronics, Год журнала: 2024, Номер 13(16), С. 3306 - 3306
Опубликована: Авг. 20, 2024
Automatic radiology report generation is a task that combines artificial intelligence and medical information processing, it fully relies on computer vision natural language processing techniques. Nowadays, automatic still very challenging because requires semantically adequate alignment of data from two modalities: images text. Existing approaches tend to focus coarse-grained at the global level do not take into account disease characteristics fine-grained semantics, which results in generated reports potentially omitting key diagnostic descriptions. In this work, we propose new approach, disease-knowledge-enhanced image–text for (DKA-RG). The method disease-level alignment, thus facilitating extraction features by model. Our approach also introduces knowledge graph inject domain expertise proposed DKA-RG consists training steps: image–report stage image-to-report stage. stage, use contrastive learning align texts high augment with enhance detection capability. text more accurate describing thanks sufficient alignment. Through extensive quantitative qualitative experiments widely used datasets, validate effectiveness our generation. achieves superior performance multiple types metrics (natural clinical efficacy metrics) compared existing methods, demonstrating can improve reliability accuracy systems.
Язык: Английский
Процитировано
0Multimedia Tools and Applications, Год журнала: 2024, Номер unknown
Опубликована: Сен. 16, 2024
Язык: Английский
Процитировано
0