Fuli Wu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Fuli Wu has a total of 87 co-authors. Here are the top co-authors by number of shared publications:
Pengyi Hao
7 shared publications- Pengyi Hao
Zhejiang University of Technology, Hangzhou, China
6 shared publications Bin Feng
Department of Oral and Maxillofacial Radiology, Stomatology Hospital, School of Stomatology, Zhejiang Provincial Clinical Research Center for Oral Diseases, Key Laboratory of Oral Biomedical Research of Zhejiang Province, Zhejiang University School of Medicine, Hangzhou, China
3 shared publications- Cong Bai
Zhejiang University of Technology
3 shared publications - Shuang Wang
Zhejiang University of Technology, Liuhe Road No.288, Hangzhou, Zhejiang, China
3 shared publications - Fan Zhang
College of Computer Science and Technology Zhejiang University of Technology Hangzhou China
2 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 12727-12735
TCFP-Sight: Ontology-Gated DAIE Reasoning for Auditable Appreciation of Traditional Chinese Figure Painting
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 12727-12735
Open AccessTCFP-Sight: Ontology-Gated DAIE Reasoning for Auditable Appreciation of Traditional Chinese Figure Painting
- Open Access
- 21 September 2026
- Published by Elsevier in SSRN Electronic Journal
MedLoop: A Validation-Gated State Construction Framework for Reducing Hallucination and Misbinding in LLM-Assisted Clinical Ultrasound Structuring
Medical vertical-domain large language models (LLMs) increasingly convert clinical ultrasound narratives into structured records, but syntactically valid JSON can still contain unsupported facts, anatomical misbindings, and invalid state updates. We propose MedLoop, a validation-gated state-construction framework that decomposes report structuring into evidence-linked, Schema-path-anchored, locally routed slot writes followed by pre-commit validation, atomic commitment, and evidence-preserving fallback. On 330 gynecologic ultrasound cases, MedLoop was compared with Direct LLM-JSON, RTR JSON (report-level template retrieval), ER-Graph Mapping, and Self-Repair / Guardrail-Retry, and was further evaluated through component ablation, model-family robustness, transfer, report realization, exploratory physician preference, and evidence-grounded question answering. MedLoop achieved Tuple-F1 of 0.9174, Path-F1 of 0.9423, Measurement-F1 of 0.8385, Subtree Exact Match of 0.8939, and Hallucination Rate of 0.0869. Relative to ER-Graph Mapping, MedLoop reduced diagnostic error events from 795 to 294, hallucinated tuples from 185 to 57, and subject/laterality misbinding events from 105 to 44. Across model families, it improved Tuple-F1 by 0.22 and Measurement-F1 by 0.28 while reducing absolute Hallucination Rate by 0.23. These results indicate that the gains arise from how candidate writes are constrained, admitted, and audited rather than from rule tables alone: under a fixed Schema and rule environment on the same 330-case gynecologic set, a rule-only configuration reached only 0.3654 Tuple-F1 and an ungated ER-Graph baseline reached 0.7077, whereas MedLoop reached 0.9174 and remained effective across five model families. MedLoop is therefore positioned as a model-agnostic control paradigm for clinical state construction, while still requiring expert-reviewed domain configuration and prospective validation.
- 21 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMedLoop: A Validation-Gated State Construction Framework for Reducing Hallucination and Misbinding in LLM-Assisted Clinical Ultrasound Structuring
Medical vertical-domain large language models (LLMs) increasingly convert clinical ultrasound narratives into structured records, but syntactically valid JSON can still contain unsupported facts, anatomical misbindings, and invalid state updates. We propose MedLoop, a validation-gated state-construction framework that decomposes report structuring into evidence-linked, Schema-path-anchored, locally routed slot writes followed by pre-commit validation, atomic commitment, and evidence-preserving fallback. On 330 gynecologic ultrasound cases, MedLoop was compared with Direct LLM-JSON, RTR JSON (report-level template retrieval), ER-Graph Mapping, and Self-Repair / Guardrail-Retry, and was further evaluated through component ablation, model-family robustness, transfer, report realization, exploratory physician preference, and evidence-grounded question answering. MedLoop achieved Tuple-F1 of 0.9174, Path-F1 of 0.9423, Measurement-F1 of 0.8385, Subtree Exact Match of 0.8939, and Hallucination Rate of 0.0869. Relative to ER-Graph Mapping, MedLoop reduced diagnostic error events from 795 to 294, hallucinated tuples from 185 to 57, and subject/laterality misbinding events from 105 to 44. Across model families, it improved Tuple-F1 by 0.22 and Measurement-F1 by 0.28 while reducing absolute Hallucination Rate by 0.23. These results indicate that the gains arise from how candidate writes are constrained, admitted, and audited rather than from rule tables alone: under a fixed Schema and rule environment on the same 330-case gynecologic set, a rule-only configuration reached only 0.3654 Tuple-F1 and an ungated ER-Graph baseline reached 0.7077, whereas MedLoop reached 0.9174 and remained effective across five model families. MedLoop is therefore positioned as a model-agnostic control paradigm for clinical state construction, while still requiring expert-reviewed domain configuration and prospective validation.
- 18 September 2026
- Published by Elsevier in Artificial Intelligence in Medicine
- Vol. 182
CReaTe-LM: A large language model for stepwise, heuristic clinical reasoning instruction
- 18 September 2026
- Published by Elsevier in Artificial Intelligence in Medicine
- Vol. 182
CReaTe-LM: A large language model for stepwise, heuristic clinical reasoning instruction
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
WFH-BFs: Dynamic Defense against DNN Bit-Flips via a Weight Function Hierarchy
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessWFH-BFs: Dynamic Defense against DNN Bit-Flips via a Weight Function Hierarchy
- 7 July 2026
- Published by Elsevier in Neural Networks
- Vol. 205  (Pt A) , 109336
Unified knowledge distillation: Integrating offline, online, and self-distillation paradigms for image classification
- 7 July 2026
- Published by Elsevier in Neural Networks
- Vol. 205  (Pt A) , 109336
Unified knowledge distillation: Integrating offline, online, and self-distillation paradigms for image classification
- Open Access
- 16 March 2026
- Published by Springer Nature in Scientific Reports
- Vol. 16  (1)
UltraReporter for transforming spoken diagnostic cues into structured ultrasound reports with large language models
Ultrasound reporting remains a manual, time-consuming process prone to errors and variability. We present UltraReporter, a compact 8B-parameter LLM pipeline that converts real-time spoken cues into structured ultrasound reports. To overcome data scarcity, we developed a multi-agent framework for synthesizing high-quality Chinese cue-report pairs from unpaired narratives. The model was further refined through template-augmented fine-tuning and defect-oriented preference optimization to ensure institutional consistency and minimize hallucinations. Evaluated on 1,311 gold-standard cases, UltraReporter outperformed nine state-of-the-art LLMs, achieving superior clinical scores (e.g. accuracy: 4.82) and NLG metrics (BLEU-4: 89.42). In a blinded reader study, its reports surpassed chief physicians in quality across normal, common, and rare cases, and 72% of prospective reports were deemed equivalent to original ones. UltraReporter integrates seamlessly into clinical workflows, generating ready-to-use reports in 1 second, significantly reducing documentation burden and demonstrating strong potential for clinical integration.
- 16 March 2026
- Published by Springer Nature in Scientific Reports
- Vol. 16  (1)
Open AccessUltraReporter for transforming spoken diagnostic cues into structured ultrasound reports with large language models
Ultrasound reporting remains a manual, time-consuming process prone to errors and variability. We present UltraReporter, a compact 8B-parameter LLM pipeline that converts real-time spoken cues into structured ultrasound reports. To overcome data scarcity, we developed a multi-agent framework for synthesizing high-quality Chinese cue-report pairs from unpaired narratives. The model was further refined through template-augmented fine-tuning and defect-oriented preference optimization to ensure institutional consistency and minimize hallucinations. Evaluated on 1,311 gold-standard cases, UltraReporter outperformed nine state-of-the-art LLMs, achieving superior clinical scores (e.g. accuracy: 4.82) and NLG metrics (BLEU-4: 89.42). In a blinded reader study, its reports surpassed chief physicians in quality across normal, common, and rare cases, and 72% of prospective reports were deemed equivalent to original ones. UltraReporter integrates seamlessly into clinical workflows, generating ready-to-use reports in 1 second, significantly reducing documentation burden and demonstrating strong potential for clinical integration.
- Open Access
- 29 November 2025
- Published by Springer Nature in Scientific Reports
- Vol. 16  (1)
Attention-Base deep learning for 3D craniofacial soft tissue landmark detection and diagnosis in orthodontics
Accurate three-dimensional (3D) craniofacial soft tissue analysis is crucial for diagnosing malocclusion and formulating personalized orthodontic treatment plans. However, the automated localization of 3D landmarks is often hindered by complex anatomy and significant biological variability.To address this challenge, we developed an innovative two-stage attention-based deep learning framework for robust landmark detection and diagnostic classification. Our approach leverages PointTransformerV3(PTv3) as its backbone, augmented by two novel modules: a Geodesic Crop module that isolates the facial region via curvature-aware geodesic masking and a Dynamic Landmark Structure Learning module that incorporates anatomical priors to model spatial interdependencies. This integrated architecture significantly enhances localization precision and structural consistency. We evaluated performance of our approach by using three complementary metrics: mean radial error (MRE), successful detection rate (SDR) across clinically relevant thresholds (2–4 mm), and successful classification rate(SCR) for treatment difficulty. Our framework achieved state-of-the-art results, with an MRE of 2.17 ± 1.54 mm (test set) and 2.19 ± 1.60 mm (validation set), alongside high SDR values meeting clinical tolerances. Notably, the model achieved a 91.74% diagnostic accuracy in classifying orthodontic treatment difficulty, underscoring its strong potential for clinical application. Comparative analyses confirmed significant improvements over existing methods in both landmark precision and diagnostic utility. Overall, these results validate the efficacy of our two-stage framework in automating craniofacial morphology assessment. By synergistically integrating geometric cropping, attention mechanisms, and anatomical constraints, the system offers orthodontists a reliable tool to enhance diagnostic precision, optimize treatment planning, and improve outcomes in patients with malocclusion.
- 29 November 2025
- Published by Springer Nature in Scientific Reports
- Vol. 16  (1)
Open AccessAttention-Base deep learning for 3D craniofacial soft tissue landmark detection and diagnosis in orthodontics
Accurate three-dimensional (3D) craniofacial soft tissue analysis is crucial for diagnosing malocclusion and formulating personalized orthodontic treatment plans. However, the automated localization of 3D landmarks is often hindered by complex anatomy and significant biological variability.To address this challenge, we developed an innovative two-stage attention-based deep learning framework for robust landmark detection and diagnostic classification. Our approach leverages PointTransformerV3(PTv3) as its backbone, augmented by two novel modules: a Geodesic Crop module that isolates the facial region via curvature-aware geodesic masking and a Dynamic Landmark Structure Learning module that incorporates anatomical priors to model spatial interdependencies. This integrated architecture significantly enhances localization precision and structural consistency. We evaluated performance of our approach by using three complementary metrics: mean radial error (MRE), successful detection rate (SDR) across clinically relevant thresholds (2–4 mm), and successful classification rate(SCR) for treatment difficulty. Our framework achieved state-of-the-art results, with an MRE of 2.17 ± 1.54 mm (test set) and 2.19 ± 1.60 mm (validation set), alongside high SDR values meeting clinical tolerances. Notably, the model achieved a 91.74% diagnostic accuracy in classifying orthodontic treatment difficulty, underscoring its strong potential for clinical application. Comparative analyses confirmed significant improvements over existing methods in both landmark precision and diagnostic utility. Overall, these results validate the efficacy of our two-stage framework in automating craniofacial morphology assessment. By synergistically integrating geometric cropping, attention mechanisms, and anatomical constraints, the system offers orthodontists a reliable tool to enhance diagnostic precision, optimize treatment planning, and improve outcomes in patients with malocclusion.
- 9 November 2024
- Published by Springer Nature in Oral Radiology
- Vol. 41  (1) , 88-101
Automatic segmentation and visualization of cortical and marrow bone in mandibular condyle on CBCT: a preliminary exploration of clinical application
- 9 November 2024
- Published by Springer Nature in Oral Radiology
- Vol. 41  (1) , 88-101
Automatic segmentation and visualization of cortical and marrow bone in mandibular condyle on CBCT: a preliminary exploration of clinical application
- Open Access
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 3028-3036
Live on the Hump: Self Knowledge Distillation via Virtual Teacher-Students Mutual Learning
For solving the limitations of the current self knowledge distillation including never fully utilizing the knowledge of shallow exits and neglecting the impact of auxiliary exits' structure on the performance of network, a novel self knowledge distillation framework via virtual teacher-students mutual learning named LOTH is proposed in this paper. A knowledgeable virtual teacher is constructed from the rich feature maps of each exit to help the learning of each exit. Meanwhile, the logit knowledges of each exit are incorporated to guide the learning of the virtual teacher. They learn mutually through the well-designed loss in LOTH. Moreover, two kinds of auxiliary building blocks are designed to balance the efficiency and effectiveness of network. Extensive experiments with diverse backbones on CIFAR-100 and Tiny-ImageNet validate the effectiveness of LOTH, which realizes superior performance with less resource by the comparison with the state-of-the-art distillation methods. The code of LOTH is available on Github https://github.com/cloak-s/LOTH.
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 3028-3036
Open AccessLive on the Hump: Self Knowledge Distillation via Virtual Teacher-Students Mutual Learning
For solving the limitations of the current self knowledge distillation including never fully utilizing the knowledge of shallow exits and neglecting the impact of auxiliary exits' structure on the performance of network, a novel self knowledge distillation framework via virtual teacher-students mutual learning named LOTH is proposed in this paper. A knowledgeable virtual teacher is constructed from the rich feature maps of each exit to help the learning of each exit. Meanwhile, the logit knowledges of each exit are incorporated to guide the learning of the virtual teacher. They learn mutually through the well-designed loss in LOTH. Moreover, two kinds of auxiliary building blocks are designed to balance the efficiency and effectiveness of network. Extensive experiments with diverse backbones on CIFAR-100 and Tiny-ImageNet validate the effectiveness of LOTH, which realizes superior performance with less resource by the comparison with the state-of-the-art distillation methods. The code of LOTH is available on Github https://github.com/cloak-s/LOTH.