BioDP-SyNet: Biophysically Orchestrated Dual-Path Synergistic Network for Breast Tumour Segmentation

BioDP-SyNet is a novel, physics-informed deep learning framework designed for precise and robust breast tumour segmentation in Dynamic Contrast-Enhanced MRI (DCE-MRI). By integrating MRI governing equations and biophysical principles, BioDP-SyNet effectively addresses challenges such as intratumoral heterogeneity, imaging noise, and artefactual interference.

Abstract

Highly sensitive to tissue haemodynamics, dynamic contrast-enhanced MRI (DCE-MRI) increasingly functions as the modality of choice for early breast-cancer screening and diagnosis, yet intratumoral heterogeneity, imaging noise and artefactual interference—coupled with CNNs’ constrained receptive fields that impede global-structure capture—erect formidable barriers to automated, precise tumour segmentation. Addressing these challenges, this work presents the Biophysically Orchestrated Dual-Path Synergistic Network (BioDP-SyNet), wherein MRI-physical governing equations and physics-informed neural paradigms converge within parallel spatial-diffusion and frequency-fluctuation pathways.

• The spatial stream embeds an Edge-Preserving Explicit Diffusion (EPED) layer for adaptive, iterative smoothing, abating signal variance and noise.
• The frequency stream integrates FreqDualis-HoloSchrod (FDH) Attention and Laplacian-Gradient (La-Gra) Attention to emulate frequency interference and phase modulation, heightening sensitivity to textural nuance and local-frequency anomalies.

Evaluation on BreastDM and BCMedSet demonstrates BioDP-SyNet's gains in Dice, IoU and 95HD, with pronounced small-lesion detection and complex-boundary delineation.

---

Core Architecture

BioDP-SyNet employs a dual-encoder, single-decoder architecture with parallel paths that synergistically extract and enhance features from two complementary physical perspectives.BioDP-SyNet employs a dual-encoder, single-decoder architecture with parallel paths that synergistically extract and enhance features from two complementary physical perspectives.

Figure: Architectural schematic of BioDP-SyNet, detailing its dual-encoder framework where the Spatial Diffusion Path (top) and Frequency Fluctuation Path (bottom) synergize to fuse multi-scale features into a shared decoder.

1. Spatial Diffusion Path

This path focuses on suppressing image noise and smoothing the highly heterogeneous signals within tumour regions.

• EPED (Edge-Preserving Explicit Diffusion) Layer: The core module, which abstracts feature maps as manifolds and evolves them using an adaptive Partial Differential Equation (PDE). It effectively removes high-frequency noise while preserving critical tumour boundary structures via a content-aware diffusion coefficient.

2. Frequency Fluctuation Path

This path is designed to overcome the local receptive field limitations of traditional convolutions, capturing macroscopic tumour morphology and fine boundary geometry.

• FDH (FreqDualis-HoloSchrod) Attention: Inspired by the Schrödinger equation from quantum mechanics, this module operates in the frequency domain to capture global shape information and phase consistency, crucial for holistic morphological assessment.

• La-Gra (Laplacian-Gradient) Attention: Embeds fixed differential operators (Laplacian and gradient) as explicit physical priors into the attention mechanism. This enables the model to better distinguish true anatomical boundaries from artifacts and enhances its sensitivity to subtle boundary details.

---

Research Highlights

• 🧠 Dual-Physics Guided: Proposes a novel dual-physics guided network (BioDP-SyNet) for breast MRI segmentation.
• 🌀 Spatial Diffusion Denoising: The spatial stream embeds a diffusion model to abate signal noise and variance.
• 🌊 Frequency Path Enhancement: Frequency path attentions enhance global morphology and boundary sensitivity.
• 🎯 Superior Performance: Demonstrates superior segmentation of small lesions and complex boundaries.

---

Experimental Results

BioDP-SyNet was benchmarked against 13 baseline models on two challenging DCE-MRI datasets.

Qualitative Comparison

BioDP-SyNet accurately delineates tumour contours, demonstrating superior robustness compared to other models, especially in difficult cases involving small lesions (Case 5) and complex backgrounds (Case 10).

Figure: A qualitative comparison against 13 leading models on representative cases. BioDP-SyNet (last column) shows the highest consistency with the expert-annotated Ground Truth.

Quantitative Results

BioDP-SyNet achieves state-of-the-art performance across all key evaluation metrics.

On BreastDM Dataset:

Model	Dice(%) $\uparrow$	IOU(%) $\uparrow$	95HD $\downarrow$	Kappa(%) $\uparrow$	MCC (%) $\uparrow$
U-Net	70.50 ± 32.26	62.20 ± 31.61	1.94 ± 0.89	70.45 ± 32.28	71.52 ± 31.75
SegNet	76.62 ± 26.04	67.53 ± 26.23	1.87 ± 0.84	76.56 ± 26.09	77.40 ± 25.65
ENet	68.62 ± 30.07	58.70 ± 28.67	2.00 ± 0.75	68.57 ± 30.08	69.78 ± 29.49
R2U-Net	76.26 ± 28.53	68.06 ± 28.34	1.83 ± 0.87	76.21 ± 28.56	77.11 ± 28.06
UNeXt	61.04 ± 29.38	49.59 ± 29.41	2.21 ± 0.80	60.99 ± 29.42	63.15 ± 28.57
MEWUNet	77.77 ± 24.73	68.58 ± 24.91	1.82 ± 0.68	77.72 ± 24.77	78.42 ± 24.33
PAttUNet	71.84 ± 33.88	64.69 ± 33.07	1.88 ± 0.87	71.80 ± 33.90	72.74 ± 33.32
DAttUNet	72.64 ± 31.87	64.72 ± 31.15	1.91 ± 0.83	72.60 ± 31.90	73.38 ± 31.59
PolypPVT	73.69 ± 22.49	62.22 ± 22.22	2.00 ± 0.65	73.63 ± 22.52	74.51 ± 22.01
MDViT	62.53 ± 28.75	51.07 ± 27.13	2.12 ± 0.71	62.46 ± 28.77	63.98 ± 28.18
VM-UNet	54.75 ± 33.95	44.75 ± 30.55	2.28 ± 0.81	54.69 ± 33.95	56.57 ± 33.53
VMamba	49.36 ± 34.92	39.80 ± 30.71	2.39 ± 0.85	49.30 ± 34.92	51.46 ± 34.65
CCViM	65.73 ± 31.60	56.00 ± 30.39	2.05 ± 0.83	65.67 ± 31.64	67.09 ± 31.06
BioDP-Sy (Ours)	79.79 ± 24.26	71.37 ± 25.25	1.69 ± 0.73	79.75 ± 24.30	80.56 ± 23.66

On BCMedSet Dataset:

Model	Dice(%) $\uparrow$	IOU(%) $\uparrow$	95HD $\downarrow$	Kappa(%) $\uparrow$	MCC (%) $\uparrow$
U-Net	80.04 ± 25.89	72.41 ± 27.03	2.10 ± 1.16	79.91 ± 25.92	78.86 ± 27.70
SegNet	79.87 ± 25.58	72.01 ± 26.61	2.11 ± 1.48	79.70 ± 25.65	79.18 ± 26.68
ENet	75.31 ± 24.58	65.07 ± 24.20	2.31 ± 1.12	75.14 ± 24.60	74.62 ± 25.74
R2U-Net	78.65 ± 27.29	70.90 ± 27.78	2.13 ± 1.19	78.51 ± 27.31	77.98 ± 28.39
UNeXt	78.36 ± 23.60	69.02 ± 24.23	2.23 ± 1.05	78.22 ± 23.62	77.25 ± 25.35
MEWUNet	79.46 ± 27.58	72.23 ± 28.19	2.06 ± 1.18	79.33 ± 27.63	78.66 ± 28.59
PAttUNet	79.92 ± 26.77	72.62 ± 27.82	2.07 ± 1.15	79.79 ± 26.81	78.75 ± 28.43
DAttUNet	80.45 ± 25.61	72.84 ± 26.60	2.08 ± 1.15	80.31 ± 25.65	78.71 ± 28.06
PolypPVT	69.74 ± 21.89	57.03 ± 21.25	2.48 ± 0.98	69.54 ± 21.87	69.42 ± 23.83
MDViT	75.41 ± 26.37	65.90 ± 25.94	2.25 ± 1.81	75.25± 26.39	74.23 ± 27.84
VM-UNet	79.34 ± 27.44	72.07 ± 28.44	2.06 ± 1.24	79.20 ± 27.50	78.13 ± 28.97
VMamba	79.91 ± 24.05	71.30 ± 24.35	2.19 ± 1.10	79.76 ± 24.09	78.00 ± 26.89
CCViM	78.89 ± 27.40	71.40 ± 28.22	2.08 ± 1.15	78.75 ± 27.46	78.15 ± 28.28
BioDP-Sy (Ours)	81.30 ± 26.86	74.68 ± 28.00	2.04 ± 1.21	81.17 ± 26.89	79.26 ± 29.41

---

Installation & Setup

1. Clone this repository:

git clone https://github.com/IMOP-lab/BioDP-SyNet.git
cd BioDP-SyNet

2. Create and activate a Conda environment:

conda create -n biodp python=3.9
conda activate biodp

Key dependencies include torch>=2.0.0 and torchvision.

---

Usage

Dataset Preparation

• BreastDM: This is a publicly available dataset. Please download it from the official source and organize it as required.
• BCMedSet: This is a private, multi-institutional dataset and is not publicly available.

Place your datasets in a data/ directory or specify the path in the training/testing commands.

Training

python train.py --dataset BreastDM --data_path /path/to/your/dataset --epochs 50 --batch_size 4 --lr 1e-4

Testing

Download our pre-trained model weights (link to be provided) and place them in the weights/ folder.

python test.py --dataset BreastDM --data_path /path/to/your/dataset --weights ./weights/biodp_synet_best.pth

---

License

This project is licensed under the MIT License.

Acknowledgments

We thank the creators and contributors of the BreastDM dataset. We also extend our gratitude to Sir Run Run Shaw Hospital and The Second Affiliated Hospital, Zhejiang University School of Medicine for their support in collecting the BCMedSet dataset.

Question

If you have any qusetion, please contact 'gaopeng.huang@hdu.edu.cn' and 'guohui71@sxmu.edu.cn'.