IMTS: An Improved Multi-scale Deep Learning Framework for Tumor Segmentation

Bladder cancer is one of the most common malignancies worldwide, with cystoscopy being a crucial diagnostic tool for its detection and monitoring. In order to offer clinicians accurate and efficient support in the diagnosis of bladder cancer, this paper proposes an improved deep learning tumor segmentation model upon cystoscopic images, i.e., IMTS model. By leveraging the multi-scale feature extraction module, light compensation pre-processing method and several post-processing methods including hole filling and largest-connected-component selection, IMTS aims to enhance the accuracy and efficiency of bladder tumor segmentation. Experimental results show that IMTS model achieves superior segmentation accuracy, with Dice and MIoU reaching 0.9099 and 0.8361, respectively.