Learning multi-scale feature representations is essential for medical image segmentation. Most existing frameworks are based on U-shape architecture in which the high-resolution representation is recovered progressively by connecting different levels of the decoder with the low-resolution representation from the encoder. However, intrinsic defects in complementary feature fusion inhibit the U-shape from aggregating efficient global and discriminative features along object boundaries. While Transformer can help model the global features, their computation complexity limits the application in real-time medical scenarios. To address these issues, we propose a Cross-scale Fusion Network (CFNet), combining a cross-scale attention module and pyramidal module to fuse multi-stage/global context information. Specifically, we first utilize large kernel convolution to design the basic building block capable of extracting global and local information. Then, we propose a Bidirectional Atrous Spatial Pyramid Pooling (BiASPP), which employs atrous convolution in the bidirectional paths to capture various shapes and sizes of brain tumors. Furthermore, we introduce a cross-stage attention mechanism to reduce redundant information when merging features from two stages with different semantics. Extensive evaluation was performed on five medical image segmentation datasets: a 3D volumetric dataset, namely Brats benchmarks. CFNet-L achieves 85.74% and 90.98% dice score for Enhanced Tumor and Whole Tumor on Brats2018, respectively. Furthermore, our largest model CFNet-L outperformed other methods on 2D medical image. It achieved 71.95%, 82.79%, and 80.79% SE for STARE, DRIVE, and CHASEDB1, respectively. The code will be available at https://github.com/aminabenabid/CFNet
27 December 2024 in IEEE Transactions on Intelligent Transportation Systems
Learning multi-scale feature representations is essential for dense prediction tasks in autonomous driving. Most existing works are based on U-shaped architectures, where high-resolution representations are progressively recovered by connecting different levels of the decoder with low-resolution representations from the encoder. We observed that rich details from low-level representation and high semantic information from high-level representations are not fully utilized in the cross-stage fusion process. Additionally, current architectures often struggle to extract efficient discriminative feature along object boundaries. To address this issue, we propose CSNet, a generic cross-stage subtraction network that extracts spatial and semantic multi-scale representations through guided contextual feature. This approach allows fine-grained features to refine deeper layers, capturing discriminative high-resolution features while filtering out redundant information. Specifically, we introduce a cross-stage subtraction module (CSM), which consists of three sub-modules: 1) a Short Path Unit, focusing on capturing complementary adjacent information; 2) Medium Path Unit for effective middle-stages features aggregation; and 3) Long Path Unit for redundant information masking and long-range context modeling. Additionally, we propose the Semantic Guided Context Reasoning (SGCR) module to reason and model contextual relations between different subtraction units. CSNet demonstrates consistent performance gains across various semantic segmentation datasets. Our model, CSNet-M, achieves 82.2% mIoU on the Camvid dataset, while CSNet-S and CSNet-M attain 79.6% and 80.5% mIoU accuracy, respectively, on the Cityscapes dataset. These results show that the proposed CSNet has the potential for enhancing real-time semantic segmentation in autonomous driving applications, offering improved accuracy and efficiency in diverse urban scenarios. The source code for this work will be published at https://github.com/mohamedac29/CSNet.