Deep models for stroke segmentation: do complex architectures always perform better?

TL;DR

The success of nnU-Net underscores the significant impact of pre- and post-processing techniques in enhancing segmentation results, rather than solely focusing on architectural designs, and suggest that proposed complex architectures may be task-specific and simpler models with appropriate pre-/post-processing pipeline can be equally or more effective in generalization across different tasks in medical image segmentation.

Abstract

Stroke segmentation plays a crucial role in the diagnosis and treatment of stroke patients by providing spatial information about affected brain regions and the extent of damage. Segmenting stroke lesions accurately is a challenging task, given that conventional manual techniques are time consuming and prone to errors. Recently, advanced deep models have been introduced for general medical image segmentation, demonstrating promising results that surpass many state of the art networks when evaluated on specific datasets. With the advent of the vision Transformers, several models have been introduced based on them, while others have aimed to design better modules based on traditional convolutional layers to extract long-range dependencies like Transformers. The question of whether such high-level designs are necessary for all segmentation cases to achieve the best results remains unanswered. In this study, we selected four types of deep models that were recently proposed and evaluated their performance for stroke segmentation: a pure Transformer-based architecture (DAE-Former), two advanced CNN-based models (LKA and DLKA) with attention mechanisms in their design, an advanced hybrid model that incorporates CNNs with Transformers (FCT), and the well-known self-adaptive nnUNet framework with its configuration based on given data. We examined their performance on two publicly available datasets, and found that the nnUNet achieved the best results with the simplest design among all. Revealing the robustness issue of Transformers to such variabilities serves as a potential reason for their weaker performance. Furthermore, nnUNet's success underscores the significant impact of preprocessing and postprocessing techniques in enhancing segmentation results, surpassing the focus solely on architectural designs

Figures (9)

• Figure 1: nnU-Net's pipeline for designing a segmentation network for each dataset isensee2021nnu.
• Figure 2: Structure of the DAE-Former architecture azad2023dae. The central component of this architecture is the Dual Transformer, comprising two distinct attention mechanisms: efficient attention for capturing spatial information and transpose attention for capturing channel information. Additionally, the skip connection cross-attention (SCCA) module is employed to integrate information from encoder layers with features from decoder layers. This fusion process enhances extracted features by preserving the most relevant information.
• Figure 3: Illustration of the main blocks in FCT tragakis2023fully. Each block comprises two key modules: Convolutional Attention and Wide Focus. Following initial convolutional layers, data undergoes processing through Convolutional Attention, akin to conventional vision Transformers. However, instead of MLP layers serving as projection layers, three layers of depth-wise convolutions are employed. Subsequently, the output is fed into the Wide Focus module, which consists of three convolutional layers: one standard and two dilated layers with varying kernel sizes to augment the receptive field. The outputs of these layers are summed and passed through an additional convolutional layer for further processing.
• Figure 4: Illustration of the main blocks in D-LKA azad2024beyond. Following processing by a convolutional layer, the data is inputted into the large kernel attention block. This block first applies a sequence of operations: a deformable depth-wise convolutional layer, followed by a deformable depth-wise dilated convolutional layer, and finally a regular convolutional layer. The attention mechanism, implemented through multiplication, serves to suppress irrelevant information, facilitating effective learning during training. Notably, in the LKA architecture, all convolutions are non-deformable, while maintaining the same underlying structure.
• Figure 5: Illustration depicting the concept of deformable convolutions, a technique that enhances standard grid sampling positions used in regular convolutions by integrating 2D offsets. This modification enables the sampling grid to flexibly deform, with the offsets learned from preceding feature maps through additional convolutional layers. As a result, this approach conditions deformation based on input features, offering localized, dense, and adaptable adjustments.