Paper Title: Cascade CNN: a two-stage segmentation framework for efficient and accurate brain tumor segmentation in multi-modal MRI
Authors: M. Vamsikrishna, Chin-Shiuh Shieh
Corresponding Author: M. Vamsikrishna (vkmangalampalli@gmail.com ), Taiwan
Abstract
The region of a brain tumor is critical in gliomas diagnosis and treatment, which involves multi-modal MRI segmentation. While segmentation models like U-Net and nnU-Net do exist, they aren’t effective in dealing with small tumor structures or with limited computational resources in general. To address these drawbacks, we propose a Cascade CNN (C-CNN) Model. C-CNN is a two-stage model that consists of two processes: coarse segmentation and refined segmentation. CoarseNet is the first process roughly segments the tumor and localizes the Region of Interest (ROI). This is succeeded by RefineNet, which does thorough multi-class segmentation on the cropped ROI, dividing the image into edema, Whole Tumor(WT), tumor core (TC), and enhancing tumor (ET). Our sequential training and multi-modal (T1, T1ce, T2, FLAIR) MRI inputs to the model reduce false positives and improve segmentation accuracy. We implemented our approach on the BraTS 2023 dataset and achieved the following Dice scores: 89.1% for WT, 83.2% for TC, 78.3% for ET, which bested single-stage models’ results. Adaptive cropping further allows for lower computational costs, enabling the algorithm to be implemented in real-time clinical settings.
Keywords
Cascade CNN, Multi-modal MRI, Medical image analysis, CoarseNet, RefineNet
