PGDiffSeg: Prior-Guided Denoising Diffusion Model with Parameter-Shared Attention for Breast Cancer Segmentation

TL;DR

A novel PGDiffSeg (Prior-Guided Diffusion Denoising Model with Parameter-Shared Attention) that applies diffusion denoising methods to breast cancer medical image segmentation, accurately recovering the affected areas from Gaussian noise is proposed.

Abstract

Early detection through imaging and accurate diagnosis is crucial in mitigating the high mortality rate associated with breast cancer. However, locating tumors from low-resolution and high-noise medical images is extremely challenging. Therefore, this paper proposes a novel PGDiffSeg (Prior-Guided Diffusion Denoising Model with Parameter-Shared Attention) that applies diffusion denoising methods to breast cancer medical image segmentation, accurately recovering the affected areas from Gaussian noise. Firstly, we design a parallel pipeline for noise processing and semantic information processing and propose a parameter-shared attention module (PSA) in multi-layer that seamlessly integrates these two pipelines. This integration empowers PGDiffSeg to incorporate semantic details at multiple levels during the denoising process, producing highly accurate segmentation maps. Secondly, we introduce a guided strategy that leverages prior knowledge to simulate the decision-making process of medical professionals, thereby enhancing the model's ability to locate tumor positions precisely. Finally, we provide the first-ever discussion on the interpretability of the generative diffusion model in the context of breast cancer segmentation. Extensive experiments have demonstrated the superiority of our model over the current state-of-the-art approaches, confirming its effectiveness as a flexible diffusion denoising method suitable for medical image research. Our code will be publicly available later.

Figures (13)

• Figure 1: The process of PGDiffSeg. It is visualized using a viridis color mapping, where the color ranges from deep purple (representing 0) to bright yellow (representing 1). This process includes both the forward process (noising) and the reverse process (denoising). At each step, a certain amount of noise is added until the image becomes a Gaussian distribution. The model learns denoising schemes using images as conditions and generates regions of interest (ROIs).
• Figure 2: Framework of PGDiffSeg model. (a) The feature pre-extraction module utilizes the slim dense block (SDB) to extract high-level semantic features from the image and $x_t$. (b) The feature encoding module performs denoising and encodes the features of $x_t$ and the image using condition units. The features are fused using parameter-shared attention (PSA) module after each layer of branches. (c) The prior knowledge guidance module plays a crucial role in the bottleneck layer by injecting expert knowledge. (d) The bottleneck layer serves as the connection point between feature encoding and feature decoding. It receives the output from feature encoding and passes it to feature decoding. (e) The feature decoding module converts the high-level feature representations from the bottleneck into the corresponding output for the original $x_t$. The model predicts the noise added in this step to obtain $x_{t-1}$.
• Figure 3: The structure of the SDB as depicted in Fig. \ref{['fig2:Overall-view-of-our-model']}(a). Taking L=4 as an example, we showcase the flow of feature vectors between different modules. Different feature vectors are combined by addition, and for better visual effects, we do not reflect this process in the figure.
• Figure 4: Parameter-shared multi-layer cross-attention (PSA) module. After being embedded in the denoising unit and condition unit of each layer, this module receives the two intermediate results of the denoising flow and condition flow in the Feature encoder, fully fuses their related information, and then transmits them to the denoising unit and condition unit of the next layer respectively.
• Figure 5: Details of prior-supervision module. To make the output of this module consistent with the Vector space of the output of the feature encoder, we designed one block and four supervised units, respectively, corresponding to the SDB and four layers units in the feature encoder