Research map / Detection and segmentation / CNN features
NaMa: Neighbor-Aware Multi-Modal Adaptive Learning for Prostate Tumor Segmentation on Anisotropic MR Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NaMa, a two-stage Unet-based network with a modality-informativeness adaptive learning module and an activation mapping guidance module for prostate tumor segmentation on anisotropic multi-modal MRI.
Problem
Prostate tumor segmentation from multi-modal (T2WI, DWI, ADC) anisotropic MR images is hard because modality informativeness varies per subject and thick slices make inter-slice consistency difficult.
Approach
Adaptively weights and fuses modality representations per subject via inter-modality correlation and gating (MIAL), then refines low-confidence slice features using neighboring-slice features selected by Wasserstein-distance confidence and guided by activation maps (AMG), with random modality masking during training.
Key results
- On IH, NaMa achieves 63.76% DSC vs 59.06% for the next best method (CSAD).
- On PICAI, NaMa achieves 59.01% DSC vs 55.89% for CSAD.
- Ablation: MIAL adds 5.13% DSC on IH and 6.89% on PICAI; AMG reduces HD95 from 12.39 to 10.45 mm on IH.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NaMa instead of CSAD | Prostate tumor segmentation · IH | DSC: 59.06 ± 0.29 -> 63.76 ± 0.24 | Helped |
| use NaMa instead of CSAD | Prostate tumor segmentation · PICAI | DSC: 55.89 ± 0.25 -> 59.01 ± 0.17 | Helped |
| use all three modalities rather than T2WI alone (ablation) | U-Net with sequence layers · IH | DSC: 39.28 ± 0.34 -> 56.44 ± 0.16 | Helped |
| use all three modalities rather than T2WI alone (ablation) | U-Net with sequence layers · PICAI | DSC: 32.41 ± 0.32 -> 50.55 ± 0.13 | Helped |
| add MIAL (ablation) | Three-modality U-Net with sequence layers · IH | DSC: 56.44 ± 0.16 -> 61.57 ± 0.15 | Helped |
| add MIAL (ablation) | Three-modality U-Net with sequence layers · PICAI | DSC: 50.55 ± 0.13 -> 57.44 ± 0.09 | Helped |
| add AMG (ablation) | Three-modality U-Net with sequence layers · IH | DSC: 56.44 ± 0.16 -> 60.63 ± 0.11 | Helped |
| add AMG (ablation) | Three-modality U-Net with sequence layers · PICAI | DSC: 50.55 ± 0.13 -> 55.16 ± 0.14 | Helped |
| add both MIAL and AMG (ablation) | Three-modality U-Net with sequence layers · IH | DSC: 56.44 ± 0.16 -> 63.76 ± 0.24 | Helped |
| add both MIAL and AMG (ablation) | Three-modality U-Net with sequence layers · PICAI | DSC: 50.55 ± 0.13 -> 59.01 ± 0.17 | Helped |
Surprising findings
- Only about 5% of cases achieve their best performance using all three modalities, suggesting multi-modal input can hurt individual cases.
- T2WI contributes less to segmentation than DWI and ADC in the baseline, despite T2WI being the annotation modality.
Limitations
- Relies on in-house data with limited size (249 training cases) and a public dataset with a random split.
- Two-stage training with frozen backbone adds procedural complexity.
- Evaluated only on prostate tumor MRI; generalization to other anatomies not shown.
Experimental setup
Datasets: in-house multi-modal prostate tumor MRI (IH, 355 patients) PICAI (PI-CAI 2022, 215 patients)
Benchmarks: in-house prostate tumor MRI dataset (IH) PICAI
Metrics: DSC HD95 ASD mIoU
Baselines: 2D nnUNet 3D nnUNet ProCDet MFSL-Net CSAD F2Net ACMINet CAT-Net
Builds on: U-Net nnU-Net Sequence layers (video latent diffusion) Focal loss Dice loss
Closest papers by meaning
- One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor Segmentation (NeurIPS 2024 · similarity 0.86)
- Cross-Modal Interactive Perception Network with Mamba for Lung Tumor Segmentation in PET-CT Images (CVPR 2025 · similarity 0.85)
- Semantic-guided Masked Mutual Learning for Multi-modal Brain Tumor Segmentation with Arbitrary Missing Modalities (AAAI 2025 · similarity 0.85)
- Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model (CVPR 2025 · similarity 0.85)
- Modality-agnostic Domain Generalizable Medical Image Segmentation by Multi-Frequency in Multi-Scale Attention (CVPR 2024 · similarity 0.84)
- TMFormer: Token Merging Transformer for Brain Tumor Segmentation with Missing Modalities (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.