cs.CVMar 12, 2026
SaveSurg-R1: A Hierarchical Reasoning Foundation Model for Scalable and Interpretable Surgical Decision Support with Multi-Center Clinical Validation
Organizations: Global College, Shanghai Jiao Tong University, Shanghai, China · Institute for Medical Engineering & Science, Massachusetts Institute of Technology, Cambridge, MA, USA · Department of Biomedical Engineering and Department of Electrical and Computer Engineering, National University of Singapore, Singapore · ICube, University of Strasbourg, CNRS, IHU Strasbourg, France · Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China · Department of Gastrointestinal Surgery, Renji Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China · College of Electronic and Information Engineering, Tongji University, Shanghai, China · Chengdu Withai Innovations Technology Company, Chengdu, China · Division of Pancreatic Surgery, Department of General Surgery, West China Hospital of Sichuan University, Chengdu, China · Department of Surgery, The Chinese University of Hong Kong, Hong Kong SAR, China · Nvidia, US · Massachusetts General Hospital, Massachusetts, US · Department of Surgery, Duke University, Durham, NC 27707, USA · School of Automation and Intelligent Sensing, Shanghai Jiao Tong University
Abstract
Surgical scene understanding demands not only accurate predictions but also interpretable reasoning that surgeons can verify against clinical expertise. However, existing surgical vision-language models generate predictions without reasoning chains, and general-purpose reasoning models fail on compositional surgical tasks without domain-specific knowledge. We present Surg-R1, a surgical Vision-Language Model that addresses this gap through hierarchical reasoning trained via a four-stage pipeline. Our approach introduces three key contributions: (1) a three-level reasoning hierarchy decomposing surgical interpretation into perceptual grounding, relational understanding, and contextual reasoning; (2) the largest surgical chain-of-thought dataset with 320,000 reasoning pairs; and (3) a four-stage training pipeline progressing from supervised fine-tuning to group relative policy optimization and iterative self-improvement. Evaluation on SurgBench, comprising six public benchmarks and six multi-center external validation datasets from five institutions, demonstrates that Surg-R1 achieves the highest Arena Score (64.9%) on public benchmarks versus Gemini 3.0 Pro (46.1%) and GPT-5.1 (37.9%), outperforming both proprietary reasoning models and specialized surgical VLMs on the majority of tasks spanning instrument localization, triplet recognition, phase recognition, action recognition, and critical view of safety assessment, with a 15.2 percentage point improvement over the strongest surgical baseline on external validation.
Figures & tables
Figure 1: Surg-R1 framework and performance. (a) Three-level chain-of-thought reasoning. Level 1 identifies instruments and tissues; Level 2 analyzes tool-tissue-action relationships; Level 3 synthesizes contextual reasoning for phase recognition. (b) Pre-training dataset ( SurgDB-3M ) coverage across multiple organ systems and procedure types. (c) Performance on public benchmark datasets (left radar) and external validation datasets (right radar) across triplet recognition, phase recognition, action recognition, and critical view of safety (CVS) assessment. (d) Four-stage training pipeline progressing from supervised fine-tuning (Stage 1) through chain-of-thought cold-start training (Stage 2) and Group Relative Policy Optimization (Stage 3) to iterative refinement (Stage 4).
Figure 2: Pre-training data, CoT data composition, and benchmark performance. (a) Pre-training dataset coverage. SurgDB-3M spans 23 surgical datasets covering six task types including tool localization, surgical VQA, triplet recognition, CVS assessment, phase recognition, and action recognition. (b) Multi-center external validation datasets. Surg-R1 achieves the highest arena score, outperforming all baseline models including surgical-domain and proprietary VLMs. (c) Public dataset benchmark. Surg-R1 achieves the highest arena score among all tested models, including reasoning-enhanced models GPT-5.1 and Gemini-3.0-Pro. Models with chain-of-thought capability are indicated by the lightbulb icon. (d) Chain-of-thought data composition. Surg-R1-Data comprises CoT conversation pairs across seven datasets, generated through surgery-aware CoT synthesis and model self-improvement.
Figure 3: Instrument localization in robotic surgery. a. Performance comparison across models using mIoU, mAP@50, mAP@75, and COCOAP metrics. b. Example of Bipolar Forceps localization. GPT-5.1 misidentifies the target instrument, predicting a bounding box in the wrong region (1066,691)-(1709,1058). Gemini 3 correctly localizes the instrument to (326,529)-(979,1045). Surg-R1 correctly identifies the bipolar forceps through Level 1 reasoning, recognizing “two flat, parallel jaws without fenestrations” characteristic of bipolar forceps, producing bounding box (327,506)-(984,1047) that closely matches ground truth (328,531)-(987,1049).
Figure 4: Triplet recognition performance on CholecT50. a. Comparison of instrument, verb, target, and triplet accuracy/mAP across models. b. Example of triplet recognition. GPT-5.1 hallucinates an additional action, predicting “grasper retract gallbladder, hook dissect peritoneum” when only one triplet is present. Gemini 3 correctly identifies the triplet. Surg-R1 applies hierarchical reasoning. Level 1 identifies “metallic silver jaws” of the grasper and “pear-shaped form with smooth, shiny surface” of the gallbladder; Level 2 analyzes the retracting action “applying traction to elevate the gallbladder away from the liver bed,” correctly synthesizing (grasper, retract, gallbladder).
Figure 5: Action recognition on SAR-RARP50 robotic surgery dataset. a. Performance comparison using accuracy, precision, recall, and F1 score. b. Example of needle manipulation action recognition. GPT-5.1 incorrectly predicts “pushing the needle through the tissue,” misinterpreting the withdrawal phase as insertion. Gemini 3 predicts “positioning the needle tip,” failing to recognize the extraction phase. Surg-R1 correctly identifies “pulling the needle out of the tissue” through hierarchical reasoning. Level 1 identifies “robotic grasper” and “suturing needle with curved metallic body”; Level 2 recognizes “needle appears to be partially emerged from the tissue” with “thread under tension,” characteristic of the “withdrawal phase of a suture pass.”
Figure 6: Phase recognition performance on Cholec80. a. Performance comparison using accuracy, recall, precision, and Jaccard index. b. Example of phase recognition. Both GPT-5.1 and Gemini 3 incorrectly predict “Gallbladder Retraction,” confusing retraction with dissection despite visible tissue manipulation. Surg-R1 correctly identifies “Gallbladder Dissection” through three-level reasoning. Level 1 recognizes the gallbladder and grasper; Level 2 analyzes “grasper retracting the gallbladder, pulling it away from the liver bed”; Level 3 synthesizes that this action “exposes the critical view of safety,” characteristic of the dissection phase rather than simple retraction.
Figure 7: Phase recognition performance on MultiBypass140. a. Cross-procedure generalization evaluation on gastric bypass surgery. b. Example of phase recognition in Roux-en-Y gastric bypass. GPT-5.1 incorrectly predicts “jejunojejunal anastomosis,” misinterpreting the scene as bowel reconnection. Gemini 3 predicts “gastrojejunal anastomosis,” confusing stomach-to-bowel connection with pouch formation. Surg-R1 correctly identifies “gastric pouch creation” through hierarchical reasoning. Level 1 identifies “linear stapler jaw” and “visible suture bites along the gastric surface”; Level 2 recognizes “stapler jaws closed on gastric wall” with “countertraction to present tissue”; Level 3 integrates these as “formation of the proximal pouch.”
Figure 8: CVS assessment on West China Hospital clinical dataset. a. Performance showing overall accuracy and per-criterion results (C1, C2, C3). b. Example assessing hepatocystic triangle clearance. LLaVA-1.5-7B correctly answers “no.” Qwen2.5-VL-7B-Surg incorrectly answers “yes,” failing to recognize incomplete clearance. Surg-R1 correctly identifies “no” with detailed reasoning, observing “no clearly defined view of the cystic duct or artery,” noting “area around gallbladder neck is obscured by fatty tissue and reflective surfaces,” and concluding “no distinct separation that would indicate a cleared triangle.”
Figure 9: CVS assessment on Nanfang Hospital clinical dataset. a. Performance comparison across models. b. Example assessing gallbladder detachment from liver bed. Gemma3-27b incorrectly answers “yes,” misinterpreting tissue positioning. Llama-4-Scout-17B correctly answers “no.” Surg-R1 provides accurate assessment with detailed reasoning. Level 1 identifies “liver bed visible as a broad, flat, reddish-brown surface beneath the gallbladder”; observes “no evident separation plane or window behind the gallbladder”; concludes “tissue planes appear intact without a clear demarcation line,” confirming the gallbladder “has not been detached from the liver bed.”
Figure 10: Triplet recognition at Renji Hospital. a. Performance showing instrument, verb, target, and triplet mAP. b. Example of triplet recognition for cystic duct clipping. Gemma3-27B predicts “grasper grasp gallbladder,” misidentifying both the instrument and target. Llama-4-Scout-17B predicts “scissors cut cystic_artery; scissors dissect adhesion,” hallucinating multiple incorrect triplets. Surg-R1 correctly identifies “clipper clip cystic_duct” through hierarchical reasoning. Level 1 recognizes “clipper with metallic silver jaws” positioned around “tubular structure” identified as the cystic duct; Level 2 analyzes “jaws closed around the duct, applying pressure to secure the clip.”
Figure 11: Action recognition performance on CUHK external validation dataset. a. Performance comparison using accuracy, recall, precision, and F1 score. b. Example of action recognition for needle manipulation. Baseline models confuse “picking-up the needle” with visually similar suturing gestures. Surg-R1 correctly identifies the action through hierarchical reasoning. Level 1 recognizes tool types and spatial arrangement; Level 2 analyzes tool-tissue interaction patterns to distinguish the current action.
Figure 12: Phase recognition on Strasbourg clinical dataset. a. Performance comparison using accuracy, precision, recall, and Jaccard index. b. Example of phase recognition in European clinical setting. Qwen2.5-VL-7B-Surg correctly identifies “Calot Triangle Dissection,” while Llama-4-Scout-17B misclassifies it as “Gallbladder Retraction.” Surg-R1 provides correct prediction with detailed reasoning. Level 1 identifies “gallbladder with pear-shaped form” and “metallic silver grasper”; Level 2 recognizes “grasper actively retracting the gallbladder to expose the Calot’s triangle”; Level 3 concludes this represents the dissection phase where “cystic duct and artery are clearly identified before division.”
Figure 13: Phase recognition at Renji Hospital. a. Performance comparison using accuracy, precision, recall, and Jaccard index. b. Example of phase recognition in laparoscopic cholecystectomy. LLaVA-1.5-7B incorrectly predicts “Calot Triangle Dissection,” while Qwen2.5-VL-7B-Surg predicts “Gallbladder Retraction,” both misinterpreting the early procedural stage. Surg-R1 correctly identifies “Preparation” phase through hierarchical reasoning. Level 1 recognizes the gallbladder with “smooth, shiny surface” and “surrounding fatty tissue”; Level 2 notes the gallbladder “positioned prominently” being “manipulated or prepared”; Level 3 concludes this represents the preparation phase where “dissection plane is being developed.”
| Task | GPT-5.1 | Gemini-3.0-Pro | Ours |
| CholecT50 Triplet | 6.77 | 8.20 | 51.69 |
| Cholec80 Phase | 49.70 | 52.35 | 80.90 |
| MultiBypass140 Phase | 29.01 | 28.72 | 49.94 |
Table 1: Performance comparison with reasoning-enhanced models. Accuracy (%) on representative tasks. Largest gap on triplet recognition (44.9 percentage points over GPT-5.1).
Figure 14: Three-level reasoning hierarchy. Level 1: visual element identification (tools, tissues, spatial relationships). Level 2: tool-tissue-action interaction analysis and triplet recognition. Level 3: procedural workflow assessment including phase recognition, step recognition, and CVS assessment.
Figure 15: Four-stage training pipeline for Surg-R1 . Stage 1: base supervised fine-tuning establishes vision-language alignment on surgical VQA data. Stage 2: chain-of-thought cold-start training introduces hierarchical reasoning structure using surgery-aware CoT synthesis. Stage 3: GRPO-based reinforcement learning refines reasoning robustness using only outcome labels. Stage 4: iterative refinement expands reasoning coverage through rejection sampling and knowledge distillation.
| Dataset Name | Surgery Type | Download URL |
| SegCol [ 17 ] | Colonoscopy | https://www.synapse.org/Synapse:syn54124209 |
| Cholec80 [ 51 ] (Phase+Tool) | Laparoscopic Cholecystectomy | https://camma.unistra.fr/datasets/ |
| Cholec80-CVS [ 41 ] | Laparoscopic Cholecystectomy | https://camma.unistra.fr/datasets/ |
| CholecT50 [ 35 ] | Laparoscopic Cholecystectomy | https://github.com/CAMMA-public/cholect50 |
| AutoLaparo [ 59 ] (Phase) | Laparoscopic Hysterectomy | https://github.com/ziyiwangx/AutoLaparo |
| AutoLaparo [ 59 ] (Seg) | Laparoscopic Hysterectomy | https://github.com/ziyiwangx/AutoLaparo |
Table 2: Public Dataset Surgery Type and Download Links
| Dataset | GPT Generated | Surg-R1 Generated | Total Conversations |
| Cholec80 [ 51 ] + CholecT50 [ 35 ] | 9,262 | 103,492 | 112,754 |
| MultiBypass140 [ 18 ] | 6,000 | 71,479 | 77,479 |
| SAR-RARP50 [ 40 ] | 5,975 | 64,530 | 70,505 |
| Endoscape [ 30 ] | 6,000 | 37,857 | 43,857 |
| EndoVis18-VQA [ 42 ] | 11,862 | 2,244 | 14,106 |
| EndoVis17 [ 4 ] | 1,776 | 0 | 1,776 |
Table 3: Statistics of conversations generated by GPT and Surg-R1 across different datasets.
| CholecT50 Triplet Recognition | |||||||||
| Accuracy (%) | mAP (%) | ||||||||
| Model | Thinking | Inst. | Verb | Target | Triplet | Inst. | Verb | Target | Triplet |
| LLaVA-1.5-7B [ 21 ] | – | 3.69 | 9.41 | 0.23 | 0.00 | 22.48 | 14.12 | 9.41 | 2.35 |
| Phi-4-Multimodal [ 1 ] | – | 14.83 | 12.98 | 3.46 | 0.12 | 22.32 | 13.99 | 9.39 | 2.35 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | – | 9.81 | 8.02 | 5.48 | 0.46 | 22.43 | 14.05 | 9.41 | 2.35 |
| InternVL3-8B [ 69 ] | – | 51.24 | 8.94 | 26.20 | 2.08 | 22.35 | 14.01 | 9.70 | 2.39 |
Table 4: Triplet recognition performance on the CholecT50 dataset. CholecT50 comprises 50 laparoscopic cholecystectomy (gallbladder removal) procedures annotated with instrument-verb-target triplets. Accuracy and mAP are reported in percentage (%). Best results are bolded .
| Cholec80 Phase Recognition (%) | |||||
| Model | Thinking | Accuracy | Recall | Precision | Jaccard |
| LLaVA-1.5-7B [ 21 ] | – | 23.46 | 14.22 | 11.53 | 6.31 |
| Phi-4-Multimodal [ 1 ] | – | 22.45 | 14.83 | 12.16 | 5.55 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | – | 22.61 | 19.87 | 26.55 | 9.92 |
| InternVL3-8B [ 69 ] | – | 23.88 | 15.21 | 15.19 | 6.67 |
| InternVL3-78B [ 69 ] | – | 27.32 | 24.03 | 30.48 | 13.30 |
Table 5: Surgical phase recognition performance on the Cholec80 dataset. Cholec80 contains 80 laparoscopic cholecystectomy procedures with seven annotated surgical phases. All metrics are reported in percentage (%). Best results are bolded .
| RARP Action Recognition (%) | |||||
| Model | Thinking | Accuracy | Recall | Precision | Jaccard |
| LLaVA-1.5-7B [ 21 ] | – | 5.10 | 12.50 | 0.64 | 0.64 |
| Phi-4-Multimodal [ 1 ] | – | 15.10 | 12.83 | 11.41 | 3.47 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | – | 12.50 | 12.68 | 5.89 | 1.78 |
| InternVL3-8B [ 69 ] | – | 29.30 | 12.78 | 9.96 | 6.13 |
| InternVL3-78B [ 69 ] | – | 29.50 | 12.87 | 22.82 | 6.28 |
Table 6: Action recognition performance on the SAR-RARP50 dataset. SAR-RARP50 contains 50 robot-assisted radical prostatectomy procedures with fine-grained surgical action annotations. All metrics are reported in percentage (%). Best results are bolded .
| MultiBypass140 Phase Recognition (%) | |||||
| Model | Thinking | Accuracy | Recall | Precision | Jaccard |
| LLaVA-1.5-7B [ 21 ] | – | 14.30 | 4.33 | 7.66 | 2.09 |
| Phi-4-Multimodal [ 1 ] | – | 6.81 | 6.76 | 7.53 | 3.14 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | – | 10.78 | 10.44 | 10.83 | 4.05 |
| InternVL3-8B [ 69 ] | – | 8.29 | 8.02 | 8.65 | 2.46 |
| MiniCPM-V-2_6 [ 65 ] | – | 9.08 | 8.83 | 9.23 | 4.06 |
Table 7: Surgical phase recognition performance on the MultiBypass140 dataset. MultiBypass140 contains 140 laparoscopic Roux-en-Y gastric bypass procedures, enabling evaluation of cross-procedure generalization beyond cholecystectomy. All metrics are reported in percentage (%). Best results are bolded .
| Dataset / Institution | Surgery Type |
| West China Hospital Dataset | Laparoscopic Cholecystectomy |
| Nanfang Hospital (SMU) Dataset | Laparoscopic Cholecystectomy |
| Renji Hospital Dataset | Laparoscopic Cholecystectomy |
| University Hospital of Strasbourg Dataset | Laparoscopic Cholecystectomy |
| Chinese University of Hong Kong (CUHK) Dataset | Robot-Assisted Radical Prostatectomy |
Table 8: Overview of Multi-Center External Validation Datasets and Associated Surgery Types
| Model | Overall | Avg | C1 | C2 | C3 |
| LLaVA-1.5-7B [ 21 ] | 33.52 | 33.41 | 31.08 | 48.35 | 20.81 |
| Phi-4-Multimodal [ 1 ] | 78.90 | 78.69 | 73.37 | 78.06 | 84.63 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | 71.77 | 71.28 | 59.28 | 84.98 | 69.58 |
| InternVL3-8B [ 69 ] | 46.75 | 47.55 | 67.47 | 30.24 | 44.93 |
| MiniCPM-V-2_6 [ 65 ] | 38.91 | 39.09 | 43.25 | 21.94 | 52.08 |
| MiniCPM-O-2_6 [ 65 ] | 39.99 | 40.89 | 63.37 | 22.47 | 36.82 |
Table 9: Critical View of Safety (CVS) assessment on the West China Hospital external validation dataset. This dataset was collected from laparoscopic cholecystectomy procedures at West China Hospital. Accuracy (%) is reported; “Overall” denotes the accuracy of the joint CVS assessment (all three criteria evaluated together), and “Avg” denotes the mean of per-criterion accuracies. Best results are bolded .
| Model | Overall | Avg | C1 | C2 | C3 |
| LLaVA-1.5-7B [ 21 ] | 33.33 | 33.33 | 34.48 | 37.93 | 27.59 |
| Phi-4-Multimodal [ 1 ] | 45.98 | 45.98 | 68.97 | 31.03 | 37.93 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | 64.37 | 64.37 | 72.41 | 62.07 | 58.62 |
| InternVL3-8B [ 69 ] | 52.87 | 52.87 | 82.76 | 24.14 | 51.72 |
| MiniCPM-V-2_6 [ 65 ] | 36.78 | 36.78 | 37.93 | 17.24 | 55.17 |
| MiniCPM-O-2_6 [ 65 ] | 40.23 | 40.23 | 68.97 | 10.34 | 41.38 |
Table 10: Critical View of Safety (CVS) assessment on the Nanfang Hospital (SMU) external validation dataset. This dataset was collected from laparoscopic cholecystectomy procedures at Nanfang Hospital. Accuracy (%) is reported; “Overall” denotes the accuracy of the joint CVS assessment, and “Avg” denotes the mean of per-criterion accuracies. Best results are bolded .
| Model | mAP (%) | |||
| Inst. | Verb | Target | Triplet | |
| LLaVA-1.5-7B [ 21 ] | 28.23 | 20.00 | 14.22 | 5.79 |
| Phi-4-Multimodal [ 1 ] | 27.65 | 20.20 | 13.47 | 5.79 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | 28.98 | 20.45 | 13.68 | 6.05 |
| InternVL3-8B [ 69 ] | 30.54 | 21.29 | 13.28 | 5.92 |
| MiniCPM-V-2_6 [ 65 ] | 27.50 | 19.87 | 13.08 | 5.79 |
Table 11: Triplet recognition performance on the Renji Hospital external validation dataset. This dataset was collected from laparoscopic cholecystectomy procedures at Renji Hospital. mAP is reported for Instrument, Verb, Target, and the complete Triplet. Best results are bolded .
| Model | Accuracy | Precision | Recall | Jaccard |
| Llama-4-Scout-17B-16E-Instruct [ 28 ] | 31.38 | 25.64 | 23.46 | 13.41 |
| Gemma3-27B-it [ 50 ] | 23.83 | 12.64 | 17.41 | 5.75 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | 20.96 | 20.72 | 16.35 | 6.89 |
| LLaVA-1.5-7B [ 21 ] | 20.21 | 11.90 | 8.38 | 3.58 |
| InternVL3-8B [ 69 ] | 17.87 | 19.92 | 18.77 | 8.27 |
| MiniCPM-V-2_6 [ 65 ] | 14.68 | 24.54 | 16.01 | 4.85 |
Table 12: Surgical phase recognition performance on the Renji Hospital external validation dataset. This dataset was collected from laparoscopic cholecystectomy procedures at Renji Hospital. Metrics include Accuracy, Macro-Precision, Macro-Recall, and Macro-Jaccard. All values are in percentages (%). Best results are bolded .
| Model | Accuracy | Precision | Recall | Jaccard |
| Llama-4-Scout-17B-16E-Instruct [ 28 ] | 40.33 | 35.68 | 24.53 | 16.63 |
| LLaVA-1.5-7B [ 21 ] | 39.50 | 19.46 | 9.90 | 6.88 |
| InternVL3-8B [ 69 ] | 37.30 | 20.68 | 15.38 | 9.68 |
| Phi-4-Multimodal [ 1 ] | 35.50 | 19.70 | 16.21 | 9.90 |
| Gemma3-27B-it [ 50 ] | 35.33 | 25.84 | 19.46 | 10.86 |
| Qwen2.5-VL-7B-Instruct [ 6 ] | 31.03 | 40.26 | 19.97 | 8.92 |
Table 13: Surgical phase recognition performance on the University Hospital of Strasbourg external validation dataset. This dataset was collected from laparoscopic cholecystectomy procedures at the University Hospital of Strasbourg. Metrics include Accuracy, Macro-Precision, Macro-Recall, and Macro-Jaccard. All values are in percentages (%). Best results are bolded .
| Model | Accuracy | Precision | Recall | Jaccard |
| LLaVA-1.5-7B [ 21 ] | 24.81 | 17.40 | 11.79 | 4.38 |
| Qwen2.5-VL-7B-Instruct [ 6 ] | 18.96 | 5.98 | 18.96 | 4.47 |
| MiniCPM-O-2_6 [ 65 ] | 18.68 | 16.18 | 18.11 | 8.05 |
| Mistral-Small-3.1-24B-Instruct-2503 [ 29 ] | 18.40 | 13.15 | 18.40 | 7.82 |
| Phi-4-Multimodal [ 1 ] | 18.30 | 12.57 | 18.21 | 5.97 |
| Gemma3-27B-it [ 50 ] | 18.11 | 8.45 | 18.11 | 4.51 |
Table 14: Action recognition performance on the Chinese University of Hong Kong (CUHK) external validation dataset. This dataset was collected from robot-assisted radical prostatectomy procedures. Metrics include Accuracy, Precision, Recall, and Jaccard. All values are in percentages (%). Best results are bolded .