基于深度学习的多任务结直肠癌分析方法研究

Research on Multitask Colorectal Cancer Analysis Method Based on Deep Learning

针对基于图像的直肠癌人工智能诊断问题,提出了一个简单而有效的结直肠空间掩膜网络(ColosMaskNet)实例分割框架,用于在MRI图像序列中联合检测和分割直肠癌的肿瘤区域。在基础网络架构中添加了一种新颖的空间注意力引导掩膜(SAG-Mask)分支,用来将物体检测器固定到框架中。还采用了改进的网络VoVNetV2作为新的骨干网络。分别对网络框架结构进行细致解析,针对不同的基础网络进行替换与实验分析,发现所采用的网络结构能够高效、准确地检测出肿瘤区域。提出了基于无锚点的单阶段检测和分割框架,通过将空间注意力导向的掩膜分支添加到框架中,实现了直肠癌肿瘤区域的实时检测和分割。

Aiming at the image-based artificial intelligence diagnosis of rectal cancer, a simple and effective colorectal spatial mask network(Colos Mask Net) instance segmentation framework is proposed for joint detection and segmentation of rectal cancer in MRI image sequences Tumor area. A novel spatial attention-guided mask(SAG-Mask) branch was added to the basic network architecture to fix the object detector to the frame. On this basis, an improved network VoVNetV2 was also used as the new backbone network. The article analyzed the network frame structure in detail, replaced and analyzed different basic networks, and found that the adopted network structure can efficiently and accurately detect the tumor area. A single-stage detection and segmentation framework based on anchor-free points was proposed. By adding spatial attention-oriented mask branches to the framework, real-time monitoring and segmentation of rectal cancer tumor regions were realized.