https://liu-yufei.github.io/tags/sam/SAMHugo Blox Builder (https://hugoblox.com)en-usTue, 01 Sep 2026 00:00:00 +0000https://liu-yufei.github.io/media/icon_hu17033853472409137288.pnghttps://liu-yufei.github.io/tags/sam/https://liu-yufei.github.io/publication/synpo/Tue, 01 Sep 2026 00:00:00 +0000https://liu-yufei.github.io/publication/synpo/<p>SynPo addresses weak negative prompts in training-free few-shot medical image segmentation. It combines SAM and DINOv2 features to construct a synergistic confidence map, then selects informative positive and negative points through confidence-aware sampling and clustering.</p> <style> .synpo-results { --synpo-navy: #183f7a; --synpo-blue: #2e64df; --synpo-line: #d9e1ec; --synpo-soft: #f5f8fc; margin: 3rem calc(50% - 50vw); padding: 0 max(1.25rem, calc((100vw - 1180px) / 2)); color: #172033; } .dark .synpo-results { --synpo-line: #334155; --synpo-soft: #111827; color: #e5e7eb; } .synpo-results__header { display: flex; align-items: end; justify-content: space-between; gap: 1rem; margin-bottom: 1.25rem; padding-bottom: .8rem; border-bottom: 5px solid var(--synpo-navy); } .synpo-results__header h2 { margin: 0; color: var(--synpo-navy); font-size: clamp(2.2rem, 5vw, 4.6rem); line-height: .95; letter-spacing: -.04em; } .dark .synpo-results__header h2 { color: #8eb5ef; } .synpo-results__header p { max-width: 34rem; margin: 0; text-align: right; font-size: .95rem; color: #64748b; } .synpo-results__layout { display: grid; grid-template-columns: minmax(0, 1.8fr) minmax(270px, .9fr); gap: 2rem; align-items: start; } .synpo-results__main, .synpo-results__side { min-width: 0; } .synpo-results__scroll { overflow-x: auto; border-top: 1px solid var(--synpo-line); border-bottom: 1px solid var(--synpo-line); } .synpo-results table { width: 100%; margin: 0; border-collapse: collapse; font-size: .79rem; line-height: 1.25; font-variant-numeric: tabular-nums; } .synpo-results th, .synpo-results td { padding: .48rem .45rem; text-align: center; white-space: nowrap; border: 0; } .synpo-results thead tr:first-child th { padding-top: .75rem; color: var(--synpo-navy); font-weight: 800; border-bottom: 1px solid var(--synpo-line); } .dark .synpo-results thead tr:first-child th { color: #a9c8f4; } .synpo-results thead tr:last-child th { background: var(--synpo-soft); font-size: .72rem; } .synpo-results th:first-child, .synpo-results td:first-child { position: sticky; left: 0; z-index: 1; text-align: left; background: #fff; font-weight: 650; } .dark .synpo-results th:first-child, .dark .synpo-results td:first-child { background: #0f172a; } .synpo-results tbody tr + tr td { border-top: 1px solid color-mix(in srgb, var(--synpo-line) 65%, transparent); } .synpo-results tbody tr.divider td { border-top: 2px solid var(--synpo-line); } .synpo-results tbody tr.synpo-row td { background: color-mix(in srgb, #dff4e9 72%, transparent); font-weight: 800; } .dark .synpo-results tbody tr.synpo-row td { background: rgba(38, 104, 72, .28); } .synpo-results tbody tr.synpo-row td:first-child { color: #166534; } .dark .synpo-results tbody tr.synpo-row td:first-child { color: #86efac; } .synpo-results__setting { display: block; margin-top: .12rem; color: #94a3b8; font-size: .64rem; font-weight: 500; text-transform: uppercase; letter-spacing: .06em; } .synpo-results__qualitative { display: block; width: 100%; margin-top: 1.15rem; border-radius: .45rem; background: #000; } .synpo-results__caption { margin: .55rem 0 0; color: #64748b; font-size: .78rem; text-align: center; } .synpo-results__side { display: grid; gap: 1.75rem; } .synpo-results__side section { padding-top: .2rem; } .synpo-results__side h3 { margin: 0 0 .8rem; font-size: 1.12rem; line-height: 1.3; } .synpo-results__side table { font-size: .86rem; } .synpo-results__side th:first-child, .synpo-results__side td:first-child { position: static; } .synpo-results__side .gain { color: var(--synpo-blue); font-weight: 750; } .synpo-results__side .best { font-weight: 850; } .synpo-results__takeaway { margin: 1.25rem 0 0; padding: .95rem 1rem; border-left: 4px solid var(--synpo-navy); background: var(--synpo-soft); font-size: .9rem; line-height: 1.55; } @media (max-width: 900px) { .synpo-results__layout { grid-template-columns: 1fr; } .synpo-results__side { grid-template-columns: repeat(2, minmax(0, 1fr)); } .synpo-results__takeaway { grid-column: 1 / -1; } } @media (max-width: 640px) { .synpo-results__header { display: block; } .synpo-results__header p { margin-top: .65rem; text-align: left; } .synpo-results__side { grid-template-columns: 1fr; } .synpo-results table { font-size: .73rem; } } </style> <section class="synpo-results" aria-labelledby="synpo-results-title"> <header class="synpo-results__header"> <h2 id="synpo-results-title">SynPo</h2> <p>Training-free few-shot medical segmentation with confidence-map synergy and informative negative prompts</p> </header> <div class="synpo-results__layout"> <div class="synpo-results__main"> <div class="synpo-results__scroll" role="region" aria-label="Segmentation benchmark results" tabindex="0"> <table> <thead> <tr> <th rowspan="2">Method</th> <th colspan="5">CHAOS-MRI</th> <th colspan="5">Synapse-CT</th> </tr> <tr> <th>Spleen</th><th>Liver</th><th>LK</th><th>RK</th><th>Mean</th> <th>Spleen</th><th>Liver</th><th>LK</th><th>RK</th><th>Mean</th> </tr> </thead> <tbody> <tr><td>SSL-ALPNet<span class="synpo-results__setting">trained</span></td><td>67.02</td><td>73.05</td><td>73.63</td><td>78.39</td><td>73.02</td><td>60.25</td><td>73.65</td><td>63.34</td><td>54.82</td><td>63.02</td></tr> <tr><td>ADNet<span class="synpo-results__setting">trained</span></td><td>75.92</td><td>80.81</td><td>75.28</td><td>83.28</td><td>78.82</td><td>63.48</td><td>77.24</td><td>72.13</td><td>79.06</td><td>72.97</td></tr> <tr><td>Q-Net<span class="synpo-results__setting">trained</span></td><td>75.99</td><td>81.74</td><td>78.36</td><td>87.98</td><td>81.02</td><td>74.86</td><td>71.21</td><td>75.26</td><td>74.79</td><td>74.03</td></tr> <tr><td>RPT<span class="synpo-results__setting">trained</span></td><td>76.37</td><td>82.86</td><td>80.72</td><td>89.82</td><td>82.44</td><td>79.13</td><td>82.57</td><td>77.05</td><td>72.58</td><td>77.83</td></tr> <tr><td>GMRD<span class="synpo-results__setting">trained</span></td><td>76.09</td><td>81.42</td><td>83.96</td><td>90.12</td><td>82.90</td><td>78.31</td><td>79.60</td><td>81.70</td><td>74.46</td><td>78.52</td></tr> <tr class="divider"><td>PerSAM<span class="synpo-results__setting">training-free</span></td><td>69.14</td><td>42.44</td><td>64.84</td><td>71.36</td><td>61.12</td><td>65.03</td><td>65.55</td><td>58.47</td><td>60.31</td><td>62.34</td></tr> <tr><td>ProtoSAM<span class="synpo-results__setting">training-free</span></td><td>76.51</td><td>81.94</td><td>71.46</td><td>81.43</td><td>77.83</td><td>65.50</td><td>87.84</td><td>69.44</td><td>71.04</td><td>73.45</td></tr> <tr class="synpo-row"><td>SynPo (ours)<span class="synpo-results__setting">training-free</span></td><td>80.30</td><td>77.32</td><td>77.32</td><td>83.04</td><td>79.50</td><td>83.76</td><td>81.32</td><td>75.00</td><td>79.63</td><td>79.91</td></tr> </tbody> </table> </div> <img class="synpo-results__qualitative" src="qualitative-results.png" alt="Qualitative abdominal organ segmentation results across four anatomical targets"> <p class="synpo-results__caption">Qualitative results on abdominal MRI. Colored overlays show the predicted organ masks.</p> </div> <aside class="synpo-results__side" aria-label="Ablation studies"> <section> <h3>Module ablation</h3> <div class="synpo-results__scroll"> <table> <thead><tr><th>SAM</th><th>DINO</th><th>PSM</th><th>NRM</th><th>Dice</th><th>Gain</th></tr></thead> <tbody> <tr><td>✓</td><td></td><td></td><td></td><td>57.48</td><td>–</td></tr> <tr><td>✓</td><td>✓</td><td></td><td></td><td>66.70</td><td class="gain">+9.22</td></tr> <tr><td>✓</td><td>✓</td><td>✓</td><td></td><td>78.74</td><td class="gain">+12.04</td></tr> <tr><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td class="best">79.93</td><td class="gain">+1.19</td></tr> </tbody> </table> </div> </section> <section> <h3>Negative-prompt strategy on CHAOS</h3> <div class="synpo-results__scroll"> <table> <thead><tr><th>Model</th><th>Dice</th></tr></thead> <tbody> <tr><td>PerSAM</td><td>56.72</td></tr> <tr><td>+ Negative Point Selection</td><td class="best">68.90</td></tr> </tbody> </table> </div> </section> <p class="synpo-results__takeaway">SynPo achieves a mean Dice score of <strong>79.91</strong> on Synapse-CT without task-specific training. The module study shows the largest gain comes from the Point Selection Module.</p> </aside> </div> </section>