Beyond Report Imitation: Clinically Aware Multi-Image Ultrasound Report Generation from Visible Evidence

1 Introduction

• Archived key frames can leave clinically valid report statements unverifiable, making them inappropriate as image-grounded targets and potentially distorting lexical evaluation.Lexical metrics may reward plausible unsupported completion while penalizing evidence-faithful omissions.Source: §1 · p. 2
• Reliable reports should preserve visible abnormalities and view-specific attribution, qualify claims when evidence is insufficient, and avoid both diagnostic escalation and conservative templates.The target is informative reporting when evidence is present and restraint when it is absent.Source: §1 · pp. 2–3
• CAMEO frames multi-image ultrasound reporting as evidence-grounded generation under partial observability, distilling visible report points into cross-view supervision and clinical-error-oriented preferences.Training proceeds through ultrasound domain adaptation, cross-view evidence grounding, and clinically aware preference alignment.Source: §1 · p. 3
• The paper introduces visual-evidence distillation, a stage-wise alignment framework combining cross-view SFT with clinical-error preferences, and an evidence-grounded LLM-assisted protocol for evaluating informativeness, diagnostic consistency, and evidence bounds.Distillation decomposes raw reports into atomic points and retains visually supportable findings as positive supervision.Source: §1 · p. 3

2.1 Radiology Report Generation and Grounded Medical Reporting

• Radiology report-generation methods such as R2Gen and R2GenGPT use paired reports as image-grounded supervision, while grounded reporting links phrases to local evidence; CAMEO first verifies archived-view support.Only verified statements are used for positive supervision, preference construction, or evidence-grounded evaluation.Source: §2.1 · p. 3

2.2 Medical MLLMs and Ultrasound Intelligence

• Medical MLLMs such as LLaVA-Med and HuatuoGPT-Vision, and ultrasound-specific EchoVLM, provide useful priors but do not specify responses to unverifiable report content; KMVE instead aligns images with raw reports.CAMEO reconstructs visually supportable evidence and evaluates against trusted visible evidence.Source: §2.2 · pp. 3–4

2.3 Multi-image and Cross-view Reasoning

• Multi-image models including MANTIS, LLaVA-OneVision [7], and LLaVA-NeXT-Interleave support comparison, but ultrasound reporting requires attributing findings to the first view, second view, both, or neither.Stage II targets clinical evidence attribution rather than generic multi-image conversation.Source: §2.3 · p. 4

2.4 Clinical Evaluation and Preference Alignment

• Clinical metrics such as CheXbert, RadGraph, RadCliQ, and GREEN generally match reference reports, whereas this setting evaluates trusted visible evidence and separates omissions, incorrect statements, and unsupported additions.Preferences target ultrasound-specific errors and shortcuts using the same clinical error axes as evaluation.Source: §2.4 · p. 4

3.1 From Partial Observability to Clinically Aware Evidence Grounding

• CAMEO distinguishes trusted visible evidence from raw-report content that may be clinically valid but unverifiable in the archived image pair.Coverage is an audit signal, not a training objective; cases are not discarded solely for low coverage.Source: §3.1 · pp. 4–5

3.2 Multi-source Ultrasound Data Organization

• PUBMEDVISION-US supports broad ultrasound instruction, while USREPORT-DISTILLED supplies 17,670 evidence-grounded instances and USREPORT-PREF supplies 21,869 training preference pairs.USREPORT-DISTILLED includes 5,890 training cases and 1,474 test cases; USREPORT-PREF also has 1,354 validation pairs.Source: §3.2 · pp. 5–6

3.3 Trustworthy Visual Evidence Distillation

• Atomic report points are visibility-verified, reconstructed into a shared evidence template, and used to generate supervision and evaluation targets.In 301 audited annotations, teacher labels reached 78.9% accuracy and 84.8% F1; 25.2% of scorable raw-report points were judged unsupported.Source: §3.3 · p. 6

3.4 Dense Evidence-grounded Supervision Construction

• Visible points provide positive targets and partial points qualified targets; cross-view SFT and clinical-error preference pairs favor informative, evidence-bounded reports over omissions, reversals, and unsupported completion.Rejected responses include view mismatches, unsupported normal templates, and conservative shortcuts.Source: §3.4 · p. 7

3.5 Stage-wise Model Alignment

• CAMEO sequentially adapts ultrasound visual-language primitives, grounds findings across views with USREPORT-DISTILLED, then uses USREPORT-PREF to calibrate behavior against clinically plausible failures.Preference alignment aims to reduce unsupported or overly specific completion without driving empty refusals when visible evidence is present.Source: §3.5 · pp. 8–9

3.6 LLM-assisted Clinical Error Evaluation

• Clinical evaluation compares reports with trusted evidence using missed-diagnosis, misdiagnosis, and overdiagnosis errors; ClinicalScore is an expert-audited proxy, not a clinical utility function.Aggregate weights prioritize omissions, then incorrect statements, then unsupported completion; all three sub-scores are reported to expose trade-offs.Source: §3.6 · p. 9

4.1 Experimental Setup

• USREPORT-DISTILLED is the primary evidence-grounded benchmark with 1,474 test cases; lexical metrics measure agreement with distilled targets, while clinical metrics use structured evidence matching.Systems receive the same case-level multi-image inputs whenever their interfaces support them.Source: §4.1 · pp. 10–11

4.2 Main Results on Multi-Image Ultrasound Report Generation

• On USREPORT-DISTILLED, CAMEO raises BLEU-1 from 0.25 to 0.40, ROUGE-1 from 0.28 to 0.45, METEOR from 0.27 to 0.43, and ClinicalScore from 55.02 to 74.20 over EchoVLM.Table 3 reports the main benchmark comparison.Source: Table 3 · p. 11, §4.2 · p. 11
• On PUBMEDVISION-US Multi, CAMEO reaches 0.38 BLEU-1, 0.44 ROUGE-1, and 0.43 METEOR, outperforming external baselines under this text-metric-only protocol.The benchmark does not support clinical error evaluation, so this result is not a general claim about ultrasound model quality.Source: §4.2 · p. 11

4.3 Component Ablation Study

• Stage II provides the largest gain through cross-view grounding, while Stage III largely preserves text metrics and improves ClinicalScore; raw-report SFT retains unverifiable content.Distilled Report SFT improves target quality but lacks explicit paired-view attribution and underperforms Stage I + Stage II.Source: §4.3 · p. 12, Table 4 · p. 12

4.4 Clinical Error Analysis

• Stage III raises overall ClinicalScore by 2.76 points, with consistent improvements for Liver and thyroid, while Breast remains limited by subtle-finding visual recall.Breast improves mainly in MisdiagScore, while its MissScore is essentially unchanged.Source: §4.4 · p. 12, Table 5 · p. 13

4.5 Qualitative Analysis and Limitations

• In Figure 4, EchoVLM and Stage I + Stage II omit the visible L-AX nodule, while CAMEO preserves its view attribution and key descriptors.The preference construction penalizes both unsupported completion and conservative shortcuts.Source: Figure 4 · p. 13, §4.5 · p. 13
• Stage II delivers the largest report-generation improvement via cross-view supervision; Stage III reshapes inclusion, omission, and qualification, with data, alignment, and evaluation treated as coupled choices.A small wording change can alter clinical meaning through polarity, attribution, or diagnostic strength.Source: §4.5 · pp. 13–14
• The evidence is limited to archived key frames and two-view reports across three common organ categories, not complete ultrasound protocols or clinical deployment.Radiologist analyses use a single reader, and ClinicalScore remains an auxiliary proxy despite correlation with reader ratings on the audited subset.Source: §4.5 · p. 14

• Cross-View Evidence Grounding yields the largest gain, while preference alignment improves the clinical error profile.Source: §5 · p. 14
• CAMEO aligns visible evidence, supervision, generation, and evaluation through atomic-point distillation, cross-view tuning, and clinical-error-oriented preferences.Source: §5 · p. 14
• Blinded radiologist studies found image-unverifiable raw-report content and general agreement between constructed preferences and expert judgment, while CAMEO retained case-specific failures.Source: §5 · p. 14
• With one reader, archived key frames, three organ categories, and an auxiliary LLM-assisted metric, results do not establish clinical utility.Source: §5 · p. 14