Multi-Level Feature-Based CNN Model for Tumor Segmentation Algorithm

This research paper addresses challenges associated with unbalanced positive and negative samples, confusion between brain and soft tissue tumors, and significant variations in the size of mediastinal lymph nodes. The objective of this investigation is to develop an effective solution for segmenting mediastinal lymph nodes using a novel multi-level feature and context segmentation network. To overcome the issues of imbalanced samples and similarities with mediastinal organs and soft tissues, the proposed approach extracts the mediastinal space using medical priority. This extraction technique minimizes the influence of mediastinal organs, enabling accurate segmentation of lymph nodes. The network architecture incorporates a global context module, enhancing the network’s ability to classify lymph nodes and backgrounds by considering global context dependencies. To tackle the challenge of varying lymph node sizes, a feature fusion module is introduced, significantly improving the segmentation accuracy of small lymph nodes within the network. Experimental results demonstrate the effectiveness of the proposed technique, achieving a segmentation accuracy of 76.93%, recall of 79.65%, and dice score of 76.08% for mediastinal lymph nodes. The proposed method outperforms existing segmentation approaches for mediastinal lymph nodes, validating its efficacy in clinical applications.