Source-linked AI summary
Multiple Myeloma Lesion Segmentation on Whole-Body Diffusion-Weighted Imaging via Efficient Anatomical Anticipation and Multimodal Confirmation
Mengmeng Zhang, Shengqian Huang, Junde Zhou, Xiaoping Wu, Hao Luog, Jing Wanga, Yicheng Sun, Jiao Li, Haibo Zhang, Sheng Xie, Fan Wangg, Qin Wangc, Huadan Xue, Yisheng Lv, Fei-yue Wang
TL;DR
Automated MM lesion segmentation on WB-DWI is challenged by limited anatomical delineation, nonspecific marrow hyperintensity, costly ROI construction, and simplistic ADC fusion. The paper proposes ADC-based bone ROI generation without dedicated bone labels and AMU-Net for anatomy-guided multimodal segmentation. It reports the best overall performance among evaluated methods, including a mean Dice score of 76.2%.
Problem
WB-DWI lesion segmentation is limited by weak anatomical delineation and nonspecific marrow hyperintensity, while bone ROI construction and ADC fusion remain practically constrained.
Method
A two-stage framework generates bone ROIs from ADC without dedicated bone labels, then uses AMU-Net to integrate ADC for structural supplementation and lesion confirmation under ROI guidance.
Results
AMU-Net achieved the best overall performance among evaluated methods, with 76.24±0.5 Dice, 65.93±1.0 IoU, and 10.1±0.7 HD95.
Takeaways & Limitations
Annotation-efficient skeletal priors and structured multimodal integration improved MM lesion segmentation performance despite limited training data.
Takeaways & Limitations
Prior ROI acquisition remains time-consuming and sensitive to patient positioning, field-of-view, and image appearance, while related ADC-fusion methods have unclear suitability for WB-DWI MM segmentation.
Abstract
from arXiv · showhide
Whole-body diffusion-weighted imaging (WB-DWI) is widely used for multiple myeloma (MM) assessment, yet automated lesion segmentation remains challenging due to limited anatomical delineation and the low specificity of marrow hyperintensity. Existing studies have introduced bone region-of-interest (ROI) information and apparent diffusion coefficient (ADC) maps to mitigate these ambiguities, but practical limitations remain. Bone ROI construction often relies on costly manual annotation, image registration, or dedicated bone models, while ADC is usually incorporated only through simple channel fusion, limiting its ability to provide complementary structural and lesion-discriminative cues. To address these limitations, we propose a two-stage framework for MM lesion segmentation on WB-DWI. In the first stage, we train a bone ROI generation model from ADC images without dedicated bone labels, providing an efficient and practical anatomical prior for lesion analysis. In the second stage, we propose Anatomy-guided Multimodal U-Net (AMU-Net), which leverages ADC in a manner consistent with clinical lesion assessment rather than treating it as a generic auxiliary modality. Extensive experiments demonstrate the effectiveness and practicality of the proposed method. It achieves the best overall performance among the evaluated methods, with a mean Dice score of 76.2%.
1. Introduction
WB-DWI segmentation is difficult because it offers limited skeletal anatomy and nonspecific marrow hyperintensity. The proposed framework addresses these issues with annotation-efficient bone ROI generation and clinically motivated ADC integration.
- WB-DWI provides limited skeletal delineation, while marrow hyperintensity can reflect MM or benign, inflammatory, edematous, and other nonmalignant processes.
- Radiologists combine anatomical context and ADC maps to assess whether suspicious signal lies in marrow-bearing bone and reflects pathological diffusion restriction.
- Existing pipelines face costly bone ROI construction and typically use ADC through simple input-level concatenation without modeling its distinct structural and lesion-confirmation roles.
- The framework generates bone ROIs from inherently co-registered ADC images without dedicated bone labels, then uses AMU-Net for ROI-guided lesion segmentation.
2. Related work
Related work uses anatomical ROIs and ADC to improve MM lesion segmentation, but practical acquisition and fusion limitations remain for WB-DWI. The proposed framework follows clinical interpretation by combining skeletal priors with multimodal lesion assessment.
- Bone or marrow ROIs provide anatomical priors that help suppress false positives and constrain predictions to plausible marrow-bearing regions.
- Prior bone ROI methods often target CT or anatomical MRI rather than WB-DWI, whose structural depiction is less clear for lesion assessment.
- Existing ROI acquisition can be time-consuming and sensitive to patient positioning, field-of-view, and image appearance, limiting practical robustness.
- ADC maps are aligned with WB-DWI and provide quantitative diffusion information that can complement assessment of pathological restriction versus normal marrow heterogeneity.
- The proposed architecture generates ADC-based bone ROI maps and combines ROI-gated ADC–DWI interaction with decoder-level lesion confirmation.
3. Method
The method uses a two-stage framework that first generates aligned soft bone ROI priors from ADC images, then performs ROI-guided multimodal lesion segmentation with ADC–DWI interactions at encoder and decoder levels.
- Framework overview: A two-stage framework first generates bone ROI maps from ADC images, then segments lesions with ADC-guided AMU-Net on WB-DWI.The stages are trained separately, with fixed Stage-I ROI probability maps supplied to Stage II.
- Stage I: Annotation-efficient bone ROI generation: Diffuse-case lesion annotations provide surrogate marrow supervision for an ADC-based ROI generator without dedicated bone labels.The generator targets a soft ROI prior with sufficient skeletal coverage rather than precise anatomical bone segmentation.
- Stage I: Annotation-efficient bone ROI generation: Voxel-wise ensemble averaging produces a more stable soft anatomical prior, and ADC–DWI inherent co-registration avoids additional registration.The ROI map is directly aligned with the lesion-segmentation input because ADC and DWI come from the same diffusion sequence.
- Stage II: Encoder-level ROI-gated cross-attention: ROI-gated cross-attention injects ADC structural cues into DWI features at the first three encoder levels while preserving DWI as the primary lesion representation.The ROI gate restricts cross-modal interaction to likely skeletal regions, reducing interference from extra-skeletal tissue.
- Stage II: Decoder-level asymmetric gate: A decoder-level asymmetric gate distinguishes concordant from discrepant DWI–ADC evidence to refine lesion confirmation under explicit ROI guidance.The agree branch can enhance or suppress embeddings, whereas the disagree branch provides one-sided suppression for potentially misleading cross-modality discrepancy.
- Stage II: Decoder-level asymmetric gate: The ADC confidence term restricts attention modulation to locations with valid ADC support before the refined decoder embedding produces lesion logits.The lesion objective combines Dice and binary crossentropy losses, with both balancing weights set to 1 unless otherwise specified.
4.1. Dataset and Data Analysis
The study uses internal and external whole-body MRI cohorts and characterizes lesion burden, sparsity, and heterogeneous multimodal tissue signals. These analyses establish the data setting and segmentation challenges.
- 67 internal cases were split into 47 development cases and 20 independent internal-validation cases, while 29 external cases were reserved for validation.
- Lesion burden was quantified as manual lesion volume, while lesion sparsity was defined as lesion-positive voxels divided by body voxels.
- Point-based analysis sampled DWI and ADC values from lesion, marrow, and background regions across 247 radiologist-annotated points from eight cases.
- The data exhibit large inter-patient variation in lesion extent, extremely sparse lesion occupancy, and heterogeneous multimodal signal characteristics.
4.2. Experimental Setup
Experiments use patient-level cross-validation, independent external testing, and multiple segmentation metrics to assess performance and generalization. The evaluation includes comparisons and ablations of the proposed Stage-II components.
- Patient-level 5-fold cross-validation assigns all records from each patient to one fold, preventing information leakage between training and testing.
- ADC distributions are examined both after pooling cases and within individual cases to assess whether absolute values or patient-specific marrow context better supports interpretation.
- All methods use identical data splits, with results reported as five-fold means and standard deviations and external testing performed independently.
- Segmentation is evaluated using Dice, IoU, HD95, precision, and recall, covering overlap, boundary accuracy, and false-positive versus false-negative trade-offs.
- The ablation analysis evaluates the encoder-level ROI-gated cross-attention and decoder-level asymmetric gate components.
4.3. Main Results
AMU-Net produces anatomically coherent ADC-based bone ROIs and achieves the strongest reported internal and external segmentation performance. Qualitative comparisons show more complete lesion coverage and fewer obvious misses.
- ADC-based ROI generation learned anatomically coherent skeletal structures and covered most lesion regions in both focal and diffuse cases despite sparse annotations.
- AMU-Net achieved the best internal performance, with Dice 76.24±0.5, IoU 65.93±1.0, and HD95 10.1±0.7.
- AMU-Net improved Dice by 3.1% and reduced HD95 relative to nnU-Net, the strongest competing baseline.
- The model achieved the second-best precision and tied for second-best recall, providing a balance between lesion sensitivity and false-positive control.
- AMU-Net produced more complete lesion coverage and fewer obvious misses than compared methods in representative qualitative cases.
- The direct ROI method better aligned with lesion-bearing skeletal regions than the registration-based pipeline, which often showed spatial mismatch.
- On the independent external cohort, AMU-Net remained best in Dice and IoU while maintaining competitive precision and recall under zero-shot testing.
4.4. Ablation Studies
Ablation studies show that both ROI-guided Stage-II components contribute to segmentation, with encoder-level ADC integration producing the larger performance drop when removed. ROI guidance supports both structural supplementation and lesion confirmation.
- Removing either the encoder-level ROI-gated cross-attention or decoder-level asymmetric gate degraded performance, while removing both produced the worst Dice and IoU.
- The larger performance drop after removing the encoder module indicates its particularly important role in supplementing structural information missing from WB-DWI.
- The decoder-level asymmetric gate further improves lesion confirmation through DWI–ADC concordance modeling.
- Removing encoder-side ROI gating or decoder-side ROI amplification reduced performance, indicating that explicit skeletal priors benefit both stages.
- Because absolute ADC values vary across patients, ROI guidance supports lesion assessment relative to each patient’s own skeletal marrow background.
4.5. Practical Value of Direct Bone ROI Generation
Direct ADC-based bone ROI generation was substantially faster and more lesion-relevant than Dixon-to-DWI registration, improving efficiency, overlap, recall, and case-level reliability.
- 4.2 s versus 43.5 s reduced per-case ROI acquisition time by nearly 10× compared with the registration-based reference.The comparison used a Dixon-to-DWI registration pipeline as a strong reference.
- Dice: 62.34 versus 23.04 improved bone–lesion overlap by 39.3 percentage points over registration.
- 70.9% versus 22.6% improved lesion recall within the ROI by 48.3 percentage points.
- 79.1% versus 0% of cases achieved Dice > 0.5, while complete failures fell from 3.0% to none.
- The ADC-based method was faster, more accurate, and more practical than registration-based ROI generation for whole-body DWI analysis.
5. Conclusion
The paper presents a two-stage framework that generates bone ROI guidance directly from ADC and uses AMU-Net for clinically motivated multimodal lesion segmentation. Despite limited training data, the method achieves clear and consistent performance gains through bone-aware guidance and structured multimodal integration.
- The two-stage framework generates bone ROI guidance directly from ADC and performs lesion segmentation with AMU-Net.
- AMU-Net leverages clinically motivated DWI–ADC interaction for structural supplementation and lesion confirmation.
- Despite limited training data, the method achieves clear and consistent performance gains for MM lesion segmentation.
- The framework supports combining bone-aware guidance with structured multimodal integration for MM lesion segmentation.
CRediT authorship contribution statement
The contribution statement assigns study conceptualization, methodology, software, drafting, data curation, validation, review, resources, supervision, and funding roles across the listed authors.
- Mengmeng Zhang contributed conceptualization, methodology, software, and original-draft writing.
- Shengqian Huang, Junde Zhou, Jing Wang, Yicheng Sun, Jiao Li, Haibo Zhang, and Sheng Xie contributed to data curation, with Huang also contributing validation.
- Xiaoping Wu contributed validation, while Hao Luo contributed review and editing.
- Fan Wang contributed resources, and Qin Wang and Huadan Xue contributed resources, supervision, and funding acquisition.