Abstract
Medical image segmentation is a critical prerequisite for quantitative analysis in clinical disease treatment. In cervical cancer radiotherapy, rapid and accurate segmentation of organs and tumors from subabdominal MRI images is essential for optimizing clinical workflows. Traditional manual annotation by specialists is time-consuming and labor-intensive. This paper proposes MFP-Net, an automatic organ segmentation algorithm that addresses the limitations of the standard U-Net architecture by incorporating dilated convolutions and a multi-scale feature pyramid structure to improve accuracy and computational efficiency.
Introduction
In the context of cervical cancer radiotherapy, precision in segmenting the bladder, rectum, and anal canal from subabdominal MRI scans is vital for effective treatment planning. While the U-Net architecture has served as the industry standard for medical image segmentation since 2015, it faces two primary challenges: the semantic gap and spatial misalignment.
The semantic gap occurs because features in the decoder often lose critical information compared to their corresponding encoder layers, leading to a mismatch in feature representation. Spatial misalignment arises from repeated convolution and pooling operations, which can blur the precise boundaries of small, irregularly shaped structures. This research introduces the Multi-scale Feature Pyramid Network (MFP-Net) to bridge these gaps while maintaining a low computational footprint.
Methodology
The MFP-Net architecture introduces several key innovations to the traditional encoder-decoder framework:
1. Dilated Convolutions
To expand the receptive field without sacrificing spatial resolution through aggressive pooling, the model utilizes dilated (atrous) convolutions. This allows the network to capture broader contextual information, which is essential for distinguishing between adjacent soft tissues in the pelvic region.
2. Feature Multiplexing and Summation
Unlike the standard U-Net, which uses "splicing" (concatenation) to join encoder and decoder features, MFP-Net employs a point-by-point sum operation. The network extracts features at multiple scales using varying dilation rates and fuses them through summation. This strategy reduces the total parameter count and helps mitigate the risk of gradient disappearance.
3. Multi-scale Feature Pyramid (MFP-Net2)
The advanced iteration of the model, MFP-Net2, utilizes multiple spatial reuse structures to form a pyramid. This ensures that high-resolution spatial relationships are transmitted directly to lower-resolution layers, assisting the decoder in recovering fine-grained spatial details lost during downsampling.
Experimental Setup
Dataset
The study utilized a clinical dataset consisting of 55 subabdominal MRI cases, totaling 3,245 slices. The target organs for segmentation included the anal canal, bladder, rectum, and the left/right femoral heads.
Preprocessing and Training
Images were scaled to 384×384 and cropped to 256×384 to manage memory consumption. To address the relatively small sample size, data augmentation techniques such as random cropping and small rotations were applied to enhance model robustness.
Results and Discussion
The performance of MFP-Net1 and MFP-Net2 was evaluated against several U-Net variants and the Transformer-based MedT model.
Segmentation Accuracy
MFP-Net demonstrated superior performance in segmenting small and irregular organs:
- Anal Canal: MFP-Net1 achieved the highest Dice Coefficient (DSC), outperforming U-Net16 by 0.031.
- Bladder and Femoral Head: MFP-Net2 yielded the best results for the bladder and left femoral head, exceeding U-Net performance by 0.017 and 0.021, respectively.
Efficiency and Model Size
MFP-Net proved to be significantly leaner than traditional models. By replacing concatenation with summation in jump connections, the authors reduced the parameter count, resulting in a smaller model size than U-Net16 without compromising accuracy.
Comparison with Transformers
Interestingly, the Transformer-based MedT model performed the worst in terms of DSC and required five times longer to train. The high video memory requirements of Transformers make them less practical for clinical environments with limited hardware, whereas MFP-Net provides a high-performance, resource-efficient alternative.
Conclusion
MFP-Net effectively addresses the semantic and spatial challenges of medical image segmentation through multi-scale feature reuse and dilated convolutions. By outperforming both traditional CNNs and modern Transformer architectures in accuracy and efficiency, it offers a viable solution for automated radiotherapy planning. Future work will explore refining multi-scale feature fusion and extending the architecture to fully 3D segmentation models.
Summary updated