M3D-Net: Hierarchical Coordination of Spatial Context, Feature Reuse, and Differential Attention for Mammography Classification

1. INTRODUCTION

• Downsampling reduces direct access to earlier patterns that must be interpreted within broader tissue structure, motivating coordinated context aggregation, feature reuse, and global mixing.Prior mammographic deep-learning work includes screening features and evaluation in routine implementation.Source: §1 · p. 1
• Prior mammography backbones emphasize different design axes, while recent work explores region-guided token selection and pretrained-model contrastive learning [10].M3D-Net aims to preserve fine-grained local cues while integrating broader tissue context as resolution decreases.Source: §1 · p. 1
• M3D-Net assigns complementary roles to multi-scale coordinate attention, multiway dynamic dense reuse, and differential attention.MCA weights local–global outputs, MUDD retrieves within-stage history before mixing, and DA subtracts global maps at coarse resolution.Source: §1 · p. 1
• The design combines explicit 2-D coordinates, bounded within-stage channel-path history, and a logit-scaled second attention map through resolution-aware placement.It builds on coordinate attention, dense connections, MUDDFormer, and Differential Transformer.Source: §1 · p. 1
• Across stages, DA is restricted to coarse resolutions, while MCA and MUDD follow stage-specific activation schedules; enabled MUDD precedes mixing and MCA follows it.The contributions also include operator analysis of constant-value rejection alongside residual routes, mammography-backbone comparison, and an adapted image–clinical ultrasound evaluation.Source: §1 · p. 1

2. M3D-NET ARCHITECTURE

• The encoder uses a stride-four stem and four resolution-and-width stages, with blocks combining positional encoding, local–global mixing, and a GN–FFN module.Residual shortcuts bypass the positional encoding, mixer, and FFN; stage and block indices determine operator activation.Source: §2 · p. 2, Figure 2 · p. 3

2.1. Context: multi-scale coordinate attention for local-global integration

• MCA divides features into channel groups, pools them at 1, 2, and 4 grids, and uses coordinate-conditioned weights to aggregate context on the input grid.The pooled groups are transformed and bilinearly upsampled before projection and normalization.Source: §2.1 · p. 2

2.2. Multiway dynamic dense connections

• MUDD retrieves at most four earlier features from the current stage through input-dependent weights, then sends the concatenated paths into the mixer.History is stored chronologically, resets at stage transitions, and retrieval is skipped when history is empty.Source: §2.2 · p. 2

2.3. Differential global attention

• Differential attention splits global queries, keys, and values into two groups, subtracts the resulting attention maps, and applies the difference to the values.The learned per-head scale starts at 0.5; with attention dropout disabled, the operator maps spatially constant values to zero before output projection, which does not imply full-encoder invariance.Source: §2.3 · pp. 2–3

2.4. Resolution-aware operator allocation

• Operator placement is resolution-aware: SRA serves stages 1–2, while DA is used in selected blocks of stages 3–4; bounded history stores at most four feature maps.The authors state that throughput and parameter-matched efficiency remain unmeasured.Source: §2.4 · p. 3

3.1. Datasets and inputs

• AISSLab contains 266 mammograms labeled normal, benign, or malignant for three-class classification, which is not pathology-confirmed diagnosis or prospective screening.The images came from the Kaggle release, whose source describes radiologist-derived annotations.Source: §3.1 · p. 3
• BrEaST includes one ultrasound image per patient for 256 patients, with a fixed, patient-disjoint split of 153 training, 51 validation, and 52 reserved test patients.Its binary target groups 98 malignant cases against 158 nonmalignant cases; inputs also include 22 clinical encodings, excluding BI-RADS, pathology diagnosis, and target labels.Source: §3.1 · p. 3

3.2. BrEaST adaptation and training

• The BrEaST comparison adapts a shared image–clinical fusion framework; M3D-Net† rearranges the DA output tensor and adds an STE projection shortcut, changing computation without adding learned parameters.The adaptation is neither a replication of the mammography implementation nor a test of a frozen mammography-trained checkpoint.Source: §3.2 · p. 3
• All models trained for 50 epochs with batch size 4 and seed 42; compatible official ImageNet weights initialized image branches, while all parameters were optimized.M3D loaded only matching names and shapes; repeated records were not additional patients.Source: §3.2 · pp. 3–4

3.3. Mammography comparison and trajectories

• 97.22% average validation accuracy over epochs 41–50 compares with 94.70% for TransXNet; M3D-Net first reaches 90% at epoch 30.The reported comparison tracks epochs, not wall-clock efficiency.Source: §3.3 · p. 4, Figure 3 · p. 4

3.4. BrEaST validation results

• The BrEaST endpoint lead over TransXNet is 1.96 points (41 versus 40 correct among 51 patients), while its Tail lead over RepViT is approximately 7.1 points.Baseline rankings differ across endpoint accuracy, Tail accuracy, and endpoint loss; Table 3 reports BrEaST multimodal validation metrics.Source: §3.4 · p. 4, Table 3 · p. 4
• BrEaST scores evaluate complete image–clinical systems, so image-only and clinical-only controls are needed to separate feature-source contributions.The authors call for shared patient-level splits, training-and-validation-only configuration selection, and reporting every predefined split and seed.Source: §3.4 · p. 4
• The evaluated settings are mammography and adapted image–clinical ultrasound; histopathology is presented as a prospective, unevaluated extension.The figure marks evaluated paths as solid and the prospective extension as dashed.Source: Figure 4 · p. 4

4. DISCUSSION

• Fixed-stage MCA removal, static-versus-dynamic MUDD weights and histories, and capacity-matched standard attention are proposed to test operator roles separately.The authors describe the operators as serving complementary feature-processing roles.Source: §4 · p. 4
• DA’s zero-sum map applies only to value aggregation; residual paths preserve other information, and the property does not establish localization, artifact suppression, or calibration.The BrEaST projection shortcut also prevents extending constant-value rejection to the full mixer.Source: §4 · p. 4
• Ultrasound adaptation changes modality, prediction endpoint, and clinical inputs; transfer benefits remain task-dependent, while histopathology and cross-site evaluation require appropriate prediction units and patient-disjoint testing.Histopathology is described as a potential extension, not an evaluated setting.Source: §4 · p. 4

• M3D-Net achieves the highest recorded validation accuracy among the compared mammography backbones.Source: §5 · p. 4
• Its resolution-aware hierarchy links stage-specific operator placement with structured interactions among spatial context weighting, bounded feature reuse, and differential aggregation.Source: §5 · p. 4
• The adapted ultrasound implementation extends evaluation to image–clinical fusion, while histopathology is suggested as a possible domain for task-specific adaptation.Repeated-seed, component-controlled, and independent evaluations are needed to characterize performance beyond the reported protocols.Source: §5 · p. 4

• The study reused public, de-identified AISSLab and BrEaST data under CC BY 4.0.Source: §6 · p. 5
• AISSLab’s source reports approval SUIRB-HR-2023-013 from the Al-Ma’amon Diagnostic Centre IRB.Source: §6 · p. 5
• BrEaST’s original collection was approved by the Lower Silesian Chamber of Medicine Bioethics Committee (2/BNR/2022), with consent waived for retrospective anonymized data.Source: §6 · p. 5
• No new participants were recruited, and no primary data were collected.Source: §6 · p. 5