Ancient ceramic texts contain rich craftsmanship knowledge but pose challenges for automatic question answering because of classical expressions, specialized terminology, ambiguous entity boundaries, and implicit relations. This study develops a domain-oriented question-answering workflow that combines structured information extraction, prompt-level knowledge provision, and parameter-efficient adaptation. BiLSTM-CRF is used to identify ceramic entities, while TextCNN classifies predefined relations among the recognized entities. The resulting entity-relation triples are stored in a source-linked knowledge base and serialized as textual context, whereas LoRA separately adapts ChatGLM3-6B to ceramic-domain question-answer data. A corpus containing 68,424 question-answer instances was constructed from six historical ceramic works. Under passage-grouped evaluation, the complete workflow achieved a Semantic Similarity of 69.5%, compared with 53.8% for ChatGLM3-6B, 60.7% for LoRA-only adaptation, 62.3% for BM25-RAG, and 64.6% for Dense-RAG. Entity recognition and relation extraction achieved F1-scores of 78.2% and 76.8%, respectively. Ablation and sensitivity analyses further characterize the performance changes associated with domain adaptation, structured context, auxiliary data, serialization format, and hyperparameter selection. The contribution lies in the construction and controlled evaluation of a source-linked ceramic question-answering workflow rather than in a new underlying neural architecture.
Accurate segmentation of kidney tumors is critical for clinical diagnosis and therapeutic planning. However, blurred tumor boundaries often make tumors difficult to distinguish from the surrounding tissues. The large variation in tumor size also leads to diverse shapes and spatial distributions in CT images, which makes reliable lesion localization difficult. To address these challenges, we propose MSB-Net, a unified coarse-to-fine 3D encoder-decoder framework that coordinates localization-oriented skip selection and high-resolution structural refinement within one network. The multiscale attention gate (MSAG) performs decoder-guided filtering on skip connections to reduce irrelevant background responses and improve localization across tumors of different sizes. The structure-aware boundary enhancement module (SBEM) then injects encoder-guided structural cues into the two highest-resolution decoder stages through bounded residual enhancement, without adding an extra boundary branch. In addition, a hybrid loss combines soft Dice, cross-entropy, and boundary-band binary cross-entropy terms, with the boundary term weighted by lambda bound=0.1. Experiments on KiTS23 and KiPA22 demonstrate consistent quantitative improvements. MSB-Net achieves a mean Dice score of 84.89% on KiTS23, exceeding the strongest baseline by 2.15 percentage points, and tumor Dice and surface Dice scores of 90.13% and 76.98% on KiPA22, respectively.
