Abstract

In the pursuit of advancing medical image segmentation, this study introduces a novel neural network, Efficient Deep Residual Network (EDRNet). Our approach is designed to handle different types of tumors, encompassing MRI brain tumor segmentation, three breast tumor ultrasound segmentation tasks, and cholorectal polyp segmentation from the Kvasir dataset. EDRNet incorporates advanced architectural features such as Enhanced Residual Dilated Blocks and a Deep Residual Network (DRNet), alongside an EfficientNet-based encoder, optimizing the extraction and processing of image features. A key innovation in our model is the integration of Spatial Channel Fusion Attention, with the DRNet which combines global and local feature extractors using an attention feature fusion module. Our modified Attention Feature Fusion Module (AFFM) plays a crucial role in integrating local features (lc), fused features (fuse), and global features (gf) to produce a rich, multi-scale representation for improved segmentation performance at decoder part. Furthermore, we have used the transfer learning based approach to train single model for the multiple cancer datasets with combine loss functions to efficiently train the neural networks. Specifically, in the MRI and breast cancer datasets, EDRNet has demonstrated good capability in properly segmenting all types of brain tumors with precise details, whether located on the left or right side of the brain. This design choice significantly enhances the model’s ability to differentiate low and high level features and boundaries in medical images, crucial for accurate segmentation. Quantitative evaluations demonstrate that EDRNet sets new benchmarks on all considered datasets. Notably, improvements in Intersection over Union (IoU) were recorded as a 9 % increase for the MRI brain tumor segmentation, a substantial 4% enhancement for breast cancer segmentation datasets, and a 1.5% improvement in cholorectal polyp segmentation on the Kvasir dataset. These results underline the efficacy of EDRNet in boosting accuracy and sensitivity, confirming its state-of-the-art performance in medical image segmentation.

Citations

Authors (3)