<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <author>
    <name>Yuelei</name>
  </author>
  <generator uri="https://hexo.io/">Hexo</generator>
  <id>https://www.406sys.top/</id>
  <link href="https://www.406sys.top/" rel="alternate"/>
  <link href="https://www.406sys.top/atom.xml" rel="self"/>
  <rights>All rights reserved 2026, Yuelei</rights>
  <subtitle>
    <![CDATA[STDIN | Think >> /dev/Mind]]>
  </subtitle>
  <title>Yuelei</title>
  <updated>2026-07-07T00:00:00.000Z</updated>
  <entry>
    <author>
      <name>Yuelei</name>
    </author>
    <category term="技术" scheme="https://www.406sys.top/categories/%E6%8A%80%E6%9C%AF/"/>
    <category term="测试" scheme="https://www.406sys.top/tags/%E6%B5%8B%E8%AF%95/"/>
    <content>
      <![CDATA[<p>最佳方案我建议是：</p><blockquote><p><strong>保留三区域 mask 的思想，但放弃“lesion &#x2F; margin &#x2F; background 三个均值 prototype”这个实现。改成 “lesion + distance-aware peritumoral rings + reference&#x2F;normal region”的 mask-guided fusion。</strong></p></blockquote><p>也就是：<strong>mask 不丢，prototype mean pooling 要换掉。</strong></p><h2 id="核心判断"><a href="#核心判断" class="headerlink" title="核心判断"></a>核心判断</h2><p>你现在失败的根因大概率不是“区域没用”，而是这三个问题叠加：</p><ol><li><strong>background 太泛</strong><br>文献支持的是 peritumoral &#x2F; parenchymal &#x2F; tumor-adjacent tissue，不是所有非病灶背景。</li><li><strong>margin 太粗</strong><br>高分文章反复说明 peritumoral 宽度很关键：0–3、3–6、6–9、9–12、12–15 mm 信息不同。你现在只有一层 margin。</li><li><strong>mean prototype 抹掉异质性</strong><br>Radiology &#x2F; Nat Commun 那几篇都在强调 habitat、subregion、ecological diversity、spatial interaction。你现在每个区域一个均值向量，信息损失太大。</li></ol><p>所以不要再继续调 <code>alpha_feat</code> 或 <code>layer3/layer4</code> 了。这个方向已经被诊断实验基本排除了。</p><h2 id="推荐新方案：Distance-aware-Peritumoral-Habitat-Fusion"><a href="#推荐新方案：Distance-aware-Peritumoral-Habitat-Fusion" class="headerlink" title="推荐新方案：Distance-aware Peritumoral Habitat Fusion"></a>推荐新方案：Distance-aware Peritumoral Habitat Fusion</h2><p>可以叫：</p><blockquote><p><strong>Mask-guided Peritumoral Habitat Fusion, MPHF</strong></p></blockquote><p>或者更贴 FedHKD：</p><blockquote><p><strong>FedHKD-RingHabitat</strong></p></blockquote><h3 id="1-区域重新定义"><a href="#1-区域重新定义" class="headerlink" title="1. 区域重新定义"></a>1. 区域重新定义</h3><p>把旧的：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line">lesion / margin / background</span><br></pre></td></tr></table></figure><p>改成：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line">lesion / peritumoral rings / reference tissue</span><br></pre></td></tr></table></figure><p>具体：</p><table><thead><tr><th>新区域</th><th>作用</th><th>对应文献启发</th></tr></thead><tbody><tr><td>Lesion</td><td>病灶主体，但不能只均值</td><td>Radiology ITH，病灶内部异质性</td></tr><tr><td>Peri-ring1</td><td>近周边</td><td>Braman 0–3 mm 与 TIL 相关</td></tr><tr><td>Peri-ring2</td><td>中周边</td><td>Braman 6–9 &#x2F; 9–12 mm 有分子亚型信息</td></tr><tr><td>Peri-ring3</td><td>远周边</td><td>Luo 15 mm 最优，20 mm 过大</td></tr><tr><td>Reference tissue</td><td>正常&#x2F;远端参考，不作为泛 background</td><td>Guo parenchymal &#x2F; glioma normal control</td></tr></tbody></table><p>如果没有真实 mm spacing，就用相对尺度：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br></pre></td><td class="code"><pre><span class="line">ring1 = 0–0.1D</span><br><span class="line">ring2 = 0.1–0.2D</span><br><span class="line">ring3 = 0.2–0.35D</span><br><span class="line">ring4 = 0.35–0.5D</span><br></pre></td></tr></table></figure><p>其中 D 可以用 lesion 等效直径或 bbox 长边。</p><p>关键：<strong>不要再用整张图剩余区域当 background。</strong></p><h2 id="2-特征表达方式改掉"><a href="#2-特征表达方式改掉" class="headerlink" title="2. 特征表达方式改掉"></a>2. 特征表达方式改掉</h2><p>不要：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line">mask × feature map → mean pooling → prototype</span><br></pre></td></tr></table></figure><p>改成：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line">mask × feature map → multi-stat / multi-token / habitat descriptor</span><br></pre></td></tr></table></figure><p>最小实现可以先做：</p><p>每个区域提取：</p><ul><li>mean；</li><li>std；</li><li>max；</li><li>top-k pooled feature；</li><li>region area ratio；</li><li>lesion-peri contrast；</li><li>peri-reference contrast。</li></ul><p>也就是每个区域不要只有一个均值，而是至少保留分布信息。</p><p>更强版本：</p><ul><li>在 lesion + peritumor 内做 patch&#x2F;habitat clustering；</li><li>提取 habitat proportion；</li><li>提取 region adjacency &#x2F; boundary interaction；</li><li>类似 NSCLC 那篇 MSI 思路。</li></ul><h2 id="3-模型结构建议"><a href="#3-模型结构建议" class="headerlink" title="3. 模型结构建议"></a>3. 模型结构建议</h2><p>我最推荐一个“安全”的结构：<strong>global branch + region residual branch</strong>。</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br></pre></td><td class="code"><pre><span class="line">whole image backbone → global logits</span><br><span class="line">region/ring features → region logits</span><br><span class="line">final logits = global logits + gated(region logits)</span><br></pre></td></tr></table></figure><p>这样有两个好处：</p><ol><li>如果区域没用，gate 可以学到接近 0，不会伤害 whole_image；</li><li>如果区域有用，它以 residual 的方式补充 whole_image。</li></ol><p>不要让区域分支完全替代 whole-image。你的 whole-image 已经很强，区域应该是补充，不是另起炉灶。</p><h2 id="4-实验路线"><a href="#4-实验路线" class="headerlink" title="4. 实验路线"></a>4. 实验路线</h2><p>我建议按三步走。</p><h3 id="Phase-1：不重训，先做-probe"><a href="#Phase-1：不重训，先做-probe" class="headerlink" title="Phase 1：不重训，先做 probe"></a>Phase 1：不重训，先做 probe</h3><p>用现有 checkpoint 抽特征，做后验 readout：</p><table><thead><tr><th>Probe</th><th>目的</th></tr></thead><tbody><tr><td>current 3-region prototype LR</td><td>旧方法锚点，已有 AUC 约 0.8876</td></tr><tr><td>lesion-only stats</td><td>看病灶内部特征</td></tr><tr><td>lesion + peri-rings stats</td><td>看周边环带是否有增量</td></tr><tr><td>lesion + peri + reference contrast</td><td>看正常参考是否有价值</td></tr><tr><td>habitat &#x2F; ring interaction features</td><td>看异质性和空间关系是否有增量</td></tr></tbody></table><p>如果 <code>lesion + peri-rings</code> 明显优于 current prototype，就证明：</p><blockquote><p>mask 有信息，但 prototype 表达方式错了。</p></blockquote><h3 id="Phase-2：训练轻量-Region-Residual-Fusion"><a href="#Phase-2：训练轻量-Region-Residual-Fusion" class="headerlink" title="Phase 2：训练轻量 Region Residual Fusion"></a>Phase 2：训练轻量 Region Residual Fusion</h3><p>只做 5 fold × 1 seed：</p><table><thead><tr><th>方法</th><th>目的</th></tr></thead><tbody><tr><td>whole_image</td><td>强基线</td></tr><tr><td>current lesion_margin_bg</td><td>旧三区域</td></tr><tr><td>lesion + peri-rings fusion</td><td>新方案</td></tr><tr><td>lesion + peri-rings + reference</td><td>验证 reference 是否有用</td></tr><tr><td>lesion + peri-rings + background</td><td>验证 background 是否拖后腿</td></tr></tbody></table><p>判据还是 paired val-AUC：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br></pre></td><td class="code"><pre><span class="line">new_method - whole_image &gt; 0.005</span><br><span class="line">且 positive folds ≥ 4/5</span><br></pre></td></tr></table></figure><h3 id="Phase-3：如果-Phase-2-positive，再接回-FedHKD"><a href="#Phase-3：如果-Phase-2-positive，再接回-FedHKD" class="headerlink" title="Phase 3：如果 Phase 2 positive，再接回 FedHKD"></a>Phase 3：如果 Phase 2 positive，再接回 FedHKD</h3><p>不要一开始就上 FedHKD 原型蒸馏。先证明区域融合本身有效。</p><p>如果有效，再做：</p><ul><li>region branch logits distillation；</li><li>habitat distribution distillation；</li><li>ring-level contrast distillation；</li><li>不再做旧的 class-region mean prototype L2。</li></ul><h2 id="最推荐的具体实现版本"><a href="#最推荐的具体实现版本" class="headerlink" title="最推荐的具体实现版本"></a>最推荐的具体实现版本</h2><p>如果你现在问我“最该先写哪个”，我选这个：</p><blockquote><p><strong>Lesion + 4 peritumoral rings + gated residual fusion + region statistics pooling</strong></p></blockquote><p>具体为：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br></pre></td><td class="code"><pre><span class="line">Input image</span><br><span class="line">  ├── global backbone → global logits</span><br><span class="line">  └── layer3 feature map</span><br><span class="line">        ├── lesion mask</span><br><span class="line">        ├── peri ring1</span><br><span class="line">        ├── peri ring2</span><br><span class="line">        ├── peri ring3</span><br><span class="line">        └── peri ring4</span><br><span class="line">             ↓</span><br><span class="line">        per-region mean/std/max pooling</span><br><span class="line">             ↓</span><br><span class="line">        region MLP / attention</span><br><span class="line">             ↓</span><br><span class="line">        region logits</span><br><span class="line">             ↓</span><br><span class="line">final logits = global logits + gate × region logits</span><br></pre></td></tr></table></figure><p>暂时不加 background。</p><p>如果一定要保留“三区域”叙事，可以这样包装：</p><figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br></pre></td><td class="code"><pre><span class="line">lesion region</span><br><span class="line">peritumoral region, internally split into distance-aware rings</span><br><span class="line">reference/background-parenchymal region, used only for contrast</span><br></pre></td></tr></table></figure><p>这样你没有放弃三区域 mask，但避免了旧 background 的噪声污染。</p><h2 id="论文叙事会更稳"><a href="#论文叙事会更稳" class="headerlink" title="论文叙事会更稳"></a>论文叙事会更稳</h2><p>你可以把现在 negative 结果写成：</p><blockquote><p>Current lesion&#x2F;margin&#x2F;background prototype distillation did not improve AUC over whole-image classification, suggesting that coarse region-mean prototypes are insufficient to exploit mask-guided information.</p></blockquote><p>然后接新方法：</p><blockquote><p>Inspired by prior peritumoral radiomics, habitat imaging, and intratumoral heterogeneity studies, we redesign mask usage from coarse region averaging to distance-aware peritumoral and heterogeneity-aware fusion.</p></blockquote><p>这就很顺：</p><ul><li>文献支持区域；</li><li>你的诊断证明旧 prototype 不行；</li><li>新方案针对性解决旧问题；</li><li>不是硬救失败结果，而是自然迭代。</li></ul><h2 id="一句话最终建议"><a href="#一句话最终建议" class="headerlink" title="一句话最终建议"></a>一句话最终建议</h2><p><strong>不要再优化原来的 lesion&#x2F;margin&#x2F;background prototype KD。最佳改进是：保留 mask，但把它改成 distance-aware peritumoral rings + heterogeneity&#x2F;statistics pooling + gated residual fusion；先证明区域融合能超过 whole-image，再考虑 FedHKD 蒸馏。</strong></p>]]>
    </content>
    <id>https://www.406sys.top/blog/mask-guided-peritumoral-habitat-fusion/</id>
    <link href="https://www.406sys.top/blog/mask-guided-peritumoral-habitat-fusion/"/>
    <published>2026-07-07T00:00:00.000Z</published>
    <summary>
      <![CDATA[<p>最佳方案我建议是：</p>
<blockquote>
<p><strong>保留三区域 mask 的思想，但放弃“lesion &#x2F; margin &#x2F; background 三个均值 prototype”这个实现。改成 “lesion + distance]]>
    </summary>
    <title>Distance-aware Peritumoral Habitat Fusion 改进方案</title>
    <updated>2026-07-07T00:00:00.000Z</updated>
  </entry>
</feed>
