SBIR Phase I Topic 402 - Artificial Intelligence-Aided Imaging for Cancer Prevention, Diagnosis, and Monitoring

该项目旨在开发一个可解释的、环路医生辅助的人工智能软件，该软件在磁共振成像中准确地描绘胶质瘤边界，计算体积曲线，并在纵向研究中对肿瘤生长进行统计量化。目前视觉分析和手工勾画轮廓线的临床实践 肿瘤是主观的，耗时的，而且经常是不一致的。MRIMath的可解释、值得信赖和环路医生人工智能系统的新颖性是多方面的。首先，我们介绍了一种多尺度特征提取框架，该框架使用初始模块对U-net图像的路径进行收缩和扩展 分段神经网络结构。其次，提出了一种新的基于修正的Dice相似系数的损失函数。第三，我们使用两种学习机制来训练和测试AI系统：学习分割肿瘤内结构和学习分割胶质瘤子区域。最后，我们产生热量 地图可视化人工智能提取的特征，从而为医生提供人工智能的注意模式和在人工智能决策期间触发的激活地图的视图。直观和交互的用户界面将允许医生检查轮廓结果、进行调整和批准轮廓， 可视化AI的解释和体积测量，最后回顾统计分析的结果。医生所做的任何修改都将在以后用于重新训练人工智能。

This project aims to develop an interpretable, physician-in-the-loop AI-aided software that accurately delineates glioma boundaries in MRIs, computes volumetric curves, and statistically quantifies the tumor growth in longitudinal studies. The current clinical practice of visually analyzing and manually contouring tumors is subjective, time-consuming, and often inconsistent. The novelty of MRIMath's explainable, trustworthy, and physician-in-the-loop AI system is multi-fold. First, we introduce a multi-scale feature extraction framework using the inception modules in contracting and expanding paths of the U-Net image segmentation neural network architecture. Second, we propose a new loss function based on the modified Dice similarity coefficient. Third, we train and test the AI system using two learning regimes: learning to segment intra-tumoral structures and learning to segment glioma sub-regions. Finally, we produce heat maps to visualize the features extracted by the AI, thus offering physicians a view of AI's attention patterns and activation maps that were triggered during AI's decision-making. An intuitive and interactive User Interface will allow the physician to review contouring results, make adjustments and approve contours, visualize AI's explanations and volumetric measurements, and finally review the results of the statistical analysis. Any modifications made by the physician will be used later to re-train AI.