18 September 2023 in IEEE Access
The generation of Traditional Chinese Medicine (TCM) prescription is one of the most challenging tasks in the research of intelligent TCM. Current researches usually use transfer learning methods to apply the relevant technology of text generation to this task simply and roughly. Either they need to train a model with large number of standardized dataset, or they ignore the domain knowledge and expertise of TCM. In order to solve these problems, we propose a hybrid neural network architecture for TCM prescription generation— PreGenerator . It includes a novel hierarchical retrieval mechanism, which can automatically extract prescription and herbal templates to facilitate accurate clinical prescription generation. Firstly, PreGenerator uses the Symptom-Prescription Retrieval (SHR) module to retrieve the most relevant prescriptions for a given patient’s symptoms. In order to follow the rule of compatibility of herbs, the Herb-Herb Retrieval (HHR) module is introduced to retrieve the next most relevant herb according to the conditioned generated herbs. Finally, the prescription decoder (PreD) fuses the symptom features, the retrieved prescription and herbal template features to generate the most relevant and effective Chinese medicine prescription. The validity of the model is verified by automatic evaluation and manual evaluation on the real medical case dataset. In addition, our model can recommend herbs that do not appear on the prescription label but are useful for relieving symptoms, which shows that our model can learn some interactions between herbs and symptoms. This research also lays a foundation for the future research on intelligent query and prescription generation of traditional Chinese medicine.
25 November 2024 in IEEE Journal of Biomedical and Health Informatics
Despite the similar global structures in Chest X-ray (CXR) images, the same anatomy exhibits varying appearances across images, including differences in local textures, shapes, colors, etc. Learning consistent representations for anatomical semantics through these diverse appearances poses a great challenge for self-supervised pre-training in CXR images. To address this challenge, we propose two new pre-training tasks: i nner- i mage a natomy l ocalization (IIAL) and c ross- i mage a natomy l ocalization (CIAL). Leveraging the relatively stable positions of identical anatomy across images, we utilize position information directly as supervision to learn consistent semantic representations. Specifically, IIAL adopts a coarse-to-fine heatmap localization approach to correlate anatomical semantics with positions, while CIAL leverages feature affine alignment and heatmap localization to establish a correspondence between identical anatomical semantics across varying images, despite their appearance diversity. Furthermore, we introduce a unified end-to-end pre-training framework, a natomy- a ware r epresentation l earning (AARL), integrating IIAL, CIAL, and a pixel restoration task. The advantages of AARL are: 1) preserving the appearance diversity; 2) training in a simple end-to-end way avoiding complicated preprocessing. Extensive experiments on six downstream tasks, including classification and segmentation tasks in various application scenarios, demonstrate that our AARL: 1) has more powerful representation and transferring ability; 2) is annotation-efficient, reducing the demand for labeled data; 3) improves the sensitivity to detecting various pathological and anatomical patterns.