Benchmarking Laryngeal Neoplasm Segmentation: A Multicenter Dataset and an Effective Method.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Accurate Laryngeal Neoplasm Segmentation (LNS) aids in laryngeal cancer diagnosis and prevention.
• Existing LNS research is hindered by a lack of public datasets.
• Colorectal Polyp Segmentation (CPS) shares similarities with LNS, but existing CPS methods show suboptimal performance on LNS tasks.

Purpose of the Study:

• To address the scarcity of LNS datasets by creating a comprehensive, multicenter dataset.
• To evaluate the performance of existing Colorectal Polyp Segmentation (CPS) methods on Laryngeal Neoplasm Segmentation (LNS).
• To propose a novel and effective segmentation method for Laryngeal Neoplasm Segmentation (LNS).

Main Methods:

• Creation of the MLN-Seg dataset, comprising 2,273 laryngeal images from four hospitals with pixel-wise annotations.
• Validation of 15 Colorectal Polyp Segmentation (CPS) methods on the MLN-Seg dataset.
• Development and implementation of the Scale-Sensitive Network (S2Net) with a Localization Calibration (LC) module for LNS.

Main Results:

• Existing CPS methods underperform on LNS, particularly with challenging cases involving blurry boundaries and camouflage.
• The proposed S2Net demonstrates superior learning ability and generalizability on the MLN-Seg dataset compared to other methods.
• S2Net achieves comparable performance on Colorectal Polyp Segmentation (CPS) tasks when evaluated on public datasets.

Conclusions:

• The MLN-Seg dataset provides a valuable resource for advancing Laryngeal Neoplasm Segmentation (LNS) research.
• The Scale-Sensitive Network (S2Net) offers an effective solution for accurate Laryngeal Neoplasm Segmentation (LNS), outperforming existing approaches.
• S2Net shows promise for both Laryngeal Neoplasm Segmentation (LNS) and Colorectal Polyp Segmentation (CPS) tasks.