26-7-31-1

ai-service/app/services/report_generator.py

@@ -0,0 +1,417 @@
+"""影像报告与临床决策建议生成。"""
+from __future__ import annotations
+
+import logging
+import re
+from typing import Any
+
+from app.schemas.models import (
+    DecisionRequest,
+    DecisionResponse,
+    Detection,
+    ImagingReportRequest,
+    ImagingReportResponse,
+    RiskItem,
+    SourceRef,
+)
+from app.services.llm_client import get_llm
+from app.services.rag_pipeline import get_rag
+
+logger = logging.getLogger(__name__)
+
+STUDY_LABEL = {
+    "X_RAY": "X 光",
+    "CT": "CT",
+    "MRI": "MRI",
+    "ULTRASOUND": "超声",
+}
+
+
+def build_imaging_texts(
+    study_type: str,
+    body_part: str,
+    detections: list[Detection],
+    mode: str,
+) -> tuple[str, str, str, float]:
+    """返回 findings, diagnosis, recommendations, confidence。"""
+    st = STUDY_LABEL.get(study_type.upper(), study_type)
+    part = body_part or "相关部位"
+    if not detections:
+        findings = f"{st}检查（{part}）：影像质量可评估，未见明确异常密度/信号灶。"
+        diagnosis = f"{part}{st}未见明显异常"
+        rec = "建议结合临床，必要时复查或进一步检查。"
+        return findings, diagnosis, rec, 0.82
+
+    lines = []
+    for d in detections:
+        name = d.label_zh or d.label
+        lines.append(
+            f"- 可见{name}样改变，框选区域约 ({int(d.bbox[0])},{int(d.bbox[1])})-"
+            f"({int(d.bbox[2])},{int(d.bbox[3])})，模型置信度 {d.confidence:.0%}"
+        )
+    findings = f"{st}检查（{part}）AI 辅助读片所见：\n" + "\n".join(lines)
+    top = max(detections, key=lambda x: x.confidence)
+    diagnosis = f"{part}可疑{top.label_zh or top.label}，建议专科医师复核"
+    rec = _rec_for_label(top.label)
+    conf = sum(d.confidence for d in detections) / len(detections)
+    if mode == "demo":
+        findings += "\n（演示模式：检测框由 YOLO 演示引擎生成，非临床验证模型输出）"
+    return findings, diagnosis, rec, round(min(0.98, conf), 4)
+
+
+def _rec_for_label(label: str) -> str:
+    mapping = {
+        "opacity": "建议结合血常规/炎症指标，必要时抗感染治疗并短期复查胸片。",
+        "nodule": "建议按结节指南分层管理，3 个月后复查 CT，必要时多学科会诊。",
+        "fracture": "建议骨科评估，必要时制动/固定，复查局部 X 光。",
+        "effusion": "建议评估积液性质，必要时穿刺或超声随访。",
+        "lesion": "建议结合临床与实验室检查，必要时增强扫描或专科转诊。",
+        "mass": "建议进一步定性检查，排除占位性病变，及时专科就诊。",
+        "calcification": "多为良性钙化可能，建议定期随访观察。",
+    }
+    return mapping.get(label, "建议专科医师综合临床资料判读，制定个体化方案。")
+
+
+def _normalize_multiline(text: str) -> str:
+    """把挤成一段的长文尽量拆成可读多行（句号/分号后换行，编号建议分行）。"""
+    if not text:
+        return ""
+    s = str(text).strip()
+    # 已有明显换行则只做空白整理
+    if "\n" in s and s.count("\n") >= 2:
+        return "\n".join(line.strip() for line in s.splitlines() if line.strip())
+
+    # 编号建议：1. / 1、 / （1） 前换行
+    s = re.sub(r"(?<![.\d])\s*([（(]?\d+[）).、])\s*", r"\n\1", s)
+    # 中文段落分隔：句号/分号后跟新意时换行（保留较短从句）
+    s = re.sub(r"([。；])\s*", r"\1\n", s)
+    lines = [ln.strip() for ln in s.splitlines() if ln.strip()]
+    # 合并过碎的短行（如单独标点）
+    merged: list[str] = []
+    for ln in lines:
+        if merged and len(ln) <= 2 and not re.match(r"^[（(]?\d+", ln):
+            merged[-1] = merged[-1] + ln
+        else:
+            merged.append(ln)
+    return "\n".join(merged)
+
+
+def _normalize_recommendations(text: str) -> str:
+    """建议统一为多行编号列表。"""
+    if not text:
+        return ""
+    s = str(text).strip()
+    # 已是多行编号
+    if re.search(r"(?m)^\s*[（(]?\d+[\.、）)]", s):
+        return "\n".join(ln.strip() for ln in s.splitlines() if ln.strip())
+
+    # 行内编号：1. / 1、 / （1）
+    items = re.findall(
+        r"[（(]?([1-9]\d?)[\.、）)]\s*([^（(]*?)(?=(?:[（(]?[1-9]\d?[\.、）)])|$)",
+        s,
+    )
+    cleaned = [(idx, t.strip(" ；;。 \t")) for idx, t in items if t.strip(" ；;。 \t")]
+    if len(cleaned) >= 2:
+        return "\n".join(f"{i}. {t}" for i, (_, t) in enumerate(cleaned, 1))
+
+    # 按分号切成条目
+    chunks = [c.strip(" ；;。") for c in re.split(r"[；;]", s) if c.strip(" ；;。")]
+    if len(chunks) >= 2:
+        return "\n".join(f"{i}. {c}" for i, c in enumerate(chunks, 1))
+    return s
+
+
+def make_full_report(
+    study_type: str,
+    body_part: str,
+    findings: str,
+    impression: str,
+    recommendations: str,
+    patient_summary: str = "",
+) -> str:
+    """结构化完整报告：固定四段，便于前端分段渲染。"""
+    st = STUDY_LABEL.get(study_type.upper(), study_type)
+    findings_n = _normalize_multiline(findings)
+    impression_n = _normalize_multiline(impression) or impression
+    rec_n = _normalize_recommendations(recommendations) or recommendations
+
+    header = [
+        "【影像诊断报告（AI 辅助）】",
+        f"检查类型：{st}",
+        f"检查部位：{body_part or '—'}",
+    ]
+    if patient_summary:
+        header.append(f"临床摘要：{patient_summary}")
+
+    sections = [
+        "\n".join(header),
+        "一、影像所见\n" + (findings_n or "—"),
+        "二、诊断印象\n" + (impression_n or "—"),
+        "三、建议\n" + (rec_n or "—"),
+        "四、声明\n本报告由 AI 辅助生成，仅供临床参考，需执业医师审核，不能替代正式报告。",
+    ]
+    return "\n\n".join(sections)
+
+
+def generate_imaging_report(req: ImagingReportRequest) -> ImagingReportResponse:
+    llm = get_llm()
+    if llm.enabled:
+        try:
+            det_lines = []
+            for d in req.detections:
+                name = d.label_zh or d.label
+                box = ",".join(str(int(x)) for x in d.bbox[:4]) if d.bbox else "-"
+                det_lines.append(f"{name} conf={d.confidence:.0%} box=[{box}]")
+            data = llm.chat_json(
+                [
+                    {
+                        "role": "system",
+                        "content": (
+                            "你是三甲医院影像科辅助报告生成器。"
+                            "必须只输出一个 JSON 对象（不要 markdown 代码块），字段："
+                            "findings（影像所见：多段文字，用换行分隔；先写检查方法与部位，"
+                            "再写病灶描述，再写其余部位阴性所见，勿写成一整段）、"
+                            "impression（诊断印象：1～3 句，可换行）、"
+                            "recommendations（建议：必须用换行的编号列表，如 "
+                            "'1. ...\\n2. ...\\n3. ...'，含进一步检查/随访/会诊）、"
+                            "full_report 不要输出（由系统按分段模板拼接）。"
+                            "依据 YOLO 检测结果撰写，专业简洁；"
+                            "明确写明需执业医师审核，不能替代正式报告。"
+                            "禁止编造未提供的患者检验结果。"
+                        ),
+                    },
+                    {
+                        "role": "user",
+                        "content": (
+                            f"检查类型={req.study_type}\n"
+                            f"检查部位={req.body_part or '未注明'}\n"
+                            f"患者摘要={req.patient_summary or '无'}\n"
+                            f"规则初诊={req.preliminary_diagnosis}\n"
+                            f"规则所见={req.findings}\n"
+                            f"检测列表:\n" + ("\n".join(det_lines) if det_lines else "（无检出）")
+                        ),
+                    },
+                ]
+            )
+            findings = _normalize_multiline(str(data.get("findings") or req.findings).strip())
+            impression = _normalize_multiline(
+                str(
+                    data.get("impression")
+                    or data.get("preliminary_diagnosis")
+                    or req.preliminary_diagnosis
+                ).strip()
+            )
+            rec = _normalize_recommendations(
+                str(data.get("recommendations") or "建议专科医师复核。").strip()
+            )
+            # 始终用分段模板拼完整报告，避免 LLM 输出一整段墙文本
+            full = make_full_report(
+                req.study_type, req.body_part, findings, impression, rec, req.patient_summary
+            )
+            model_name = llm.info().get("model") or "llm"
+            logger.info("影像报告已由 LLM 生成 model=%s", model_name)
+            return ImagingReportResponse(
+                findings=findings,
+                impression=impression,
+                recommendations=rec,
+                full_report=full,
+                model_version=f"report-llm:{model_name}",
+            )
+        except Exception as e:
+            logger.warning("影像报告 LLM 失败，回退模板: %s", e)
+
+    logger.info("影像报告使用模板模式（LLM 未启用或调用失败） llm_enabled=%s", llm.enabled)
+    # recommendations 留空，由 imaging API 回退到 build_imaging_texts 的按病灶建议
+    full = make_full_report(
+        req.study_type,
+        req.body_part,
+        req.findings,
+        req.preliminary_diagnosis,
+        "建议结合临床，由影像科/临床医师最终签发。",
+        req.patient_summary,
+    )
+    return ImagingReportResponse(
+        findings=req.findings,
+        impression=req.preliminary_diagnosis,
+        recommendations="",
+        full_report=full,
+        model_version="report-template",
+    )
+
+
+def generate_decision(req: DecisionRequest) -> DecisionResponse:
+    rag = get_rag()
+    query = " ".join(
+        x for x in [req.diagnosis, req.chief_complaint, req.history, req.imaging_summary] if x
+    ).strip() or "常见病辅助决策"
+    sources = rag.retrieve(query, top_k=4)
+    context = rag.build_context(sources)
+    patient = req.patient
+    patient_desc = ""
+    if patient:
+        patient_desc = f"年龄={patient.age} 性别={patient.gender} 姓名={patient.name or ''}"
+
+    llm = get_llm()
+    if llm.enabled:
+        try:
+            data = llm.chat_json(
+                [
+                    {
+                        "role": "system",
+                        "content": (
+                            "你是临床辅助决策系统。输出严格 JSON，字段："
+                            "treatment_suggestions(数组，元素含 title,description,confidence),"
+                            "medication_suggestions(数组，元素含 name,dosage,category,confidence),"
+                            "nursing_advice(字符串数组),"
+                            "follow_up_plan(字符串数组),"
+                            "risks(数组，元素含 type,description,level,confidence),"
+                            "conflicts(字符串数组),"
+                            "full_text(字符串)。"
+                            "必须提醒需医师审核；勿编造不存在的检查结果。"
+                        ),
+                    },
+                    {
+                        "role": "user",
+                        "content": (
+                            f"患者：{patient_desc}\n"
+                            f"主诉：{req.chief_complaint}\n"
+                            f"病史：{req.history}\n"
+                            f"查体：{req.exam_findings}\n"
+                            f"诊断：{req.diagnosis}\n"
+                            f"用药：{req.medications}\n"
+                            f"影像摘要：{req.imaging_summary}\n"
+                            f"知识库：\n{context or '无'}"
+                        ),
+                    },
+                ]
+            )
+            return _map_decision(data, sources, engine="fastapi-rag+llm")
+        except Exception as e:
+            logger.warning("决策 LLM 失败: %s", e)
+
+    return _template_decision(req, sources)
+
+
+def _map_decision(data: dict[str, Any], sources: list[SourceRef], engine: str) -> DecisionResponse:
+    risks = []
+    for r in data.get("risks") or []:
+        if isinstance(r, dict):
+            risks.append(
+                RiskItem(
+                    type=str(r.get("type") or "风险"),
+                    description=str(r.get("description") or ""),
+                    level=str(r.get("level") or "中"),
+                    confidence=float(r.get("confidence") or 0.8),
+                )
+            )
+    return DecisionResponse(
+        treatment_suggestions=list(data.get("treatment_suggestions") or []),
+        medication_suggestions=list(data.get("medication_suggestions") or []),
+        nursing_advice=[str(x) for x in (data.get("nursing_advice") or [])],
+        follow_up_plan=[str(x) for x in (data.get("follow_up_plan") or [])],
+        risks=risks,
+        conflicts=[str(x) for x in (data.get("conflicts") or [])],
+        sources=sources,
+        full_text=str(data.get("full_text") or ""),
+        engine=engine,
+    )
+
+
+def _template_decision(req: DecisionRequest, sources: list[SourceRef]) -> DecisionResponse:
+    dx = req.diagnosis or ""
+    treatments: list[dict[str, Any]] = []
+    meds: list[dict[str, Any]] = []
+    nursing: list[str] = []
+    follow: list[str] = []
+    risks: list[RiskItem] = []
+
+    if "高血压" in dx:
+        treatments = [
+            {"title": "生活方式干预", "description": "低盐饮食，适量有氧运动，控制体重，戒烟限酒", "confidence": 0.95},
+            {"title": "药物治疗", "description": "可考虑 ACEI/ARB 或 CCB 作为一线方案（需医师确认）", "confidence": 0.9},
+        ]
+        meds = [
+            {"name": "氨氯地平", "dosage": "5mg qd", "category": "钙通道阻滞剂", "confidence": 0.9},
+            {"name": "缬沙坦", "dosage": "80mg qd", "category": "ARB", "confidence": 0.88},
+        ]
+        nursing = ["监测血压并记录", "宣教服药依从性", "观察头晕、乏力等低血压症状"]
+        follow = ["1–2 周门诊复查血压", "评估靶器官损害相关检查"]
+    elif "糖尿病" in dx:
+        treatments = [
+            {"title": "饮食运动", "description": "控制总热量与碳水，规律运动", "confidence": 0.95},
+            {"title": "降糖治疗", "description": "二甲双胍等一线方案需结合肾功能与禁忌", "confidence": 0.9},
+        ]
+        meds = [{"name": "二甲双胍", "dosage": "0.5g tid", "category": "双胍类", "confidence": 0.92}]
+        nursing = ["血糖监测指导", "足部护理宣教", "低血糖识别与处理"]
+        follow = ["2–4 周复诊评估血糖", "定期查 HbA1c"]
+    elif "肺炎" in dx or "阴影" in dx:
+        treatments = [
+            {"title": "抗感染", "description": "根据社区/医院获得性肺炎指南选择抗生素", "confidence": 0.88},
+            {"title": "支持治疗", "description": "休息、补液、必要时氧疗", "confidence": 0.92},
+        ]
+        meds = [{"name": "阿莫西林", "dosage": "0.5g tid", "category": "青霉素类", "confidence": 0.85}]
+        nursing = ["监测体温与呼吸", "叩背排痰指导", "隔离防护宣教（如需要）"]
+        follow = ["3–5 天评估疗效", "必要时复查胸片"]
+    elif "结节" in dx:
+        treatments = [
+            {"title": "分层随访", "description": "按结节大小与特征选择随访或进一步检查", "confidence": 0.9},
+        ]
+        nursing = ["戒烟宣教", "避免焦虑，说明随访意义"]
+        follow = ["3 个月复查 CT", "出现咯血/胸痛及时就诊"]
+    else:
+        treatments = [
+            {"title": "进一步评估", "description": "完善相关检查以明确诊断", "confidence": 0.85},
+            {"title": "对症处理", "description": "根据症状给予相应支持治疗", "confidence": 0.88},
+        ]
+        nursing = ["观察病情变化", "用药与生活方式宣教"]
+        follow = ["按病情 1–2 周复诊", "出现加重症状及时急诊"]
+
+    if req.patient and req.patient.age and req.patient.age >= 65:
+        risks.append(
+            RiskItem(
+                type="高龄风险",
+                description="高龄患者需注意剂量调整、跌倒与多药联用风险",
+                level="高",
+                confidence=0.85,
+            )
+        )
+    if "高血压" in dx and req.patient and req.patient.age and req.patient.age > 60:
+        risks.append(
+            RiskItem(
+                type="心血管风险",
+                description="高血压合并高龄，心血管事件风险增加",
+                level="中",
+                confidence=0.8,
+            )
+        )
+
+    conflicts: list[str] = []
+    meds_text = req.medications or ""
+    if "华法林" in meds_text and "阿司匹林" in meds_text:
+        conflicts.append("警告：华法林与阿司匹林联合使用可能增加出血风险")
+    if "ACEI" in meds_text and "保钾" in meds_text:
+        conflicts.append("注意：ACEI 与保钾利尿剂联用可能致高钾血症")
+
+    src_hint = ""
+    if sources:
+        src_hint = "\n知识库参考：" + "；".join(s.title for s in sources[:3])
+
+    full = (
+        f"诊断相关辅助建议（规则+RAG）：{dx or '未明确'}\n"
+        f"治疗：{'; '.join(t['title'] for t in treatments)}\n"
+        f"护理：{'；'.join(nursing)}\n"
+        f"随访：{'；'.join(follow)}"
+        f"{src_hint}\n"
+        "（模板模式，可配置 LLM_API_KEY 启用大模型增强）"
+    )
+    return DecisionResponse(
+        treatment_suggestions=treatments,
+        medication_suggestions=meds,
+        nursing_advice=nursing,
+        follow_up_plan=follow,
+        risks=risks,
+        conflicts=conflicts,
+        sources=sources,
+        full_text=full,
+        engine="fastapi-rag-template",
+    )