<p>Uveal melanoma requires volumetric mapping to guide therapy, as two-dimensional frames lack the topology necessary for precise boundary delineation. This work employs DUSt3R-based correspondence estimation with self-calibrated poses and dense pointmaps, refined through intrinsic reprojection loss, while Logarithmic Positional Partition Interval Encoding (LPPIE) is applied to depth data, pointmaps, and camera parameter tuples to reduce memory usage. The pipeline was evaluated on melanoma, nevus, melanosis, pterygium, and phantom ocular images using metrics including completeness (C), mean absolute error (MAE), root mean squared error (RMSE), rotation error, and translation error. For melanoma cases, <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10278_2025_1684_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="66" /> </InlineMediaObject> <EquationSource Format="TEX">$C \approx 0.43$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>C</mi> <mo>≈</mo> <mn>0.43</mn> </mrow> </math></EquationSource> </InlineEquation>, rotation mismatch <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10278_2025_1684_Article_IEq2.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="35" /> </InlineMediaObject> <EquationSource Format="TEX">$\approx 2^{\circ }$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <msup> <mn>2</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation>, RMSE <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10278_2025_1684_Article_IEq3.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="56" /> </InlineMediaObject> <EquationSource Format="TEX">$\approx 0.021$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <mn>0.021</mn> </mrow> </math></EquationSource> </InlineEquation>, and translation error <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10278_2025_1684_Article_IEq4.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="40" /> </InlineMediaObject> <EquationSource Format="TEX">$\approx 3.9$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <mn>3.9</mn> </mrow> </math></EquationSource> </InlineEquation> cm were observed, whereas simpler morphologies achieved <i>C</i> up to 0.61 with stable MAE between 0.005 and 0.008 and <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10278_2025_1684_Article_IEq5.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="112" /> </InlineMediaObject> <EquationSource Format="TEX">$\delta < 1.25 \approx 0.98$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>δ</mi> <mo><</mo> <mn>1.25</mn> <mo>≈</mo> <mn>0.98</mn> </mrow> </math></EquationSource> </InlineEquation>–0.99. LPPIE substantially reduced the memory footprint, with only minor texture degradation near vascular branching. The method produced coherent meshes under modest computational requirements, though robustness decreased for irregular surfaces, specular reflections, and motion artifacts. The approach operates with single-camera inputs, minimal calibration, and portable hardware, suggesting its potential for accessible ocular volumetry and opportunities for refinement through region constraints, enhanced glare suppression, and precision scheduling.</p>

Pointmaps to Practice: 3D Multi-view Ocular Lesion Mapping

期刊论文
出版模式：

开放获取
发表日期：

2025年10月24日

查看全文

Vasileios Alevizos,
George A. Papakostas

摘要

Uveal melanoma requires volumetric mapping to guide therapy, as two-dimensional frames lack the topology necessary for precise boundary delineation. This work employs DUSt3R-based correspondence estimation with self-calibrated poses and dense pointmaps, refined through intrinsic reprojection loss, while Logarithmic Positional Partition Interval Encoding (LPPIE) is applied to depth data, pointmaps, and camera parameter tuples to reduce memory usage. The pipeline was evaluated on melanoma, nevus, melanosis, pterygium, and phantom ocular images using metrics including completeness (C), mean absolute error (MAE), root mean squared error (RMSE), rotation error, and translation error. For melanoma cases, $C \approx 0.43$ $C \approx 0.43$ , rotation mismatch $\approx 2^{\circ }$ $\approx 2^{\circ}$ , RMSE $\approx 0.021$ $\approx 0.021$ , and translation error $\approx 3.9$ $\approx 3.9$ cm were observed, whereas simpler morphologies achieved C up to 0.61 with stable MAE between 0.005 and 0.008 and $\delta < 1.25 \approx 0.98$ $δ < 1.25 \approx 0.98$ –0.99. LPPIE substantially reduced the memory footprint, with only minor texture degradation near vascular branching. The method produced coherent meshes under modest computational requirements, though robustness decreased for irregular surfaces, specular reflections, and motion artifacts. The approach operates with single-camera inputs, minimal calibration, and portable hardware, suggesting its potential for accessible ocular volumetry and opportunities for refinement through region constraints, enhanced glare suppression, and precision scheduling.