Excellence in Research: Harnessing Big Data and Domain Knowledge to Advance Deep Learning for Interpretable Cell Quantitation

准确、可靠和可解释的胰腺细胞识别和评估可以更好地预测身体对疾病或治疗反应的变化。因此，它对临床实践和实验室工作中的病理学家和科学家非常有价值，他们照顾患者，特别是那些患有肥胖症或糖尿病的患者。该项目设计和开发了新的深度学习、迁移学习和可解释的方法，并将其与领域专家知识相结合，通过利用和利用丰富的大数据，实现可靠、可解释和准确的胰腺细胞定量。现有的方法要么手动识别，要么需要了解大量的参数设置，以实现自动化。此外，这些技术不强加可解释性和忽视领域专家知识。该项目通过结合领域专家知识，从其他图像处理和分割问题中转移知识和经验以适应细胞定量来解决这些问题。项目成果将使生物医学成像、迁移学习、人机交互等领域的研究人员受益，并为学生提供深度学习、图像处理和可解释方法等领域的实用学习材料。此外，该项目有可能提高研究能力和合作，为来自代表性不足的社区的学生创造新的研究机会，以攻读STEM高级学位。该项目开发数据驱动的算法，解决可解释的细胞定量问题。通过新颖的深度学习算法，该项目利用大数据的潜力来获得可靠，可解释和准确的评估。此外，该项目还探索了一个通用框架，用于扩展现有的研究结果，并结合领域专家知识，以补充建模和学习过程。具体来说，这项研究使用显微图像，机器学习，迁移学习和领域知识在三个方面：（1）基于深度学习的细胞定量算法。该算法具有可解释性，以领域专家知识为指导，具有可信、可靠和准确的性能。(2)加强迁移学习领域的数据驱动技术，以填补深度学习和生物医学成像交叉领域的重要知识空白。(3)推进领域知识整合和人工干预方法，用于开发和操作基于深度学习的生物医学研究和医疗保健系统。所有的研究成果都是公开的，以促进扩展和加速不同社区的研究人员和学生的各种应用程序的开发。这个奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的知识价值和更广泛的影响审查标准进行评估的支持。

Accurate, reliable, and interpretable pancreatic cell recognition and evaluation allow better anticipation of changes in body response to disease or treatment. Thus, it is highly valuable to pathologists and scientists in clinical practice and laboratory work, who care for patients, especially for those with commodities like obesity or diabetes. This project designs and develops novel deep learning, transfer learning, and interpretable methods integrated with domain-expert knowledge for reliable, interpretable, and accurate pancreatic cell quantitation by harnessing and exploiting the abundance of big data. Existing methods do either manual recognition or require understanding of a huge number of parameters setting for automation. Additionally, these techniques do not impose interpretability and disregard domain-expert knowledge. The project tackles these issues by incorporating domain expert knowledge, transferring knowledge and experience from other image processing and segmentation problems adapting for cell quantitation. The project outcomes would benefit researchers in biomedical imaging, transfer learning, human-machine interaction in addition to providing practical studying materials in areas such as deep learning, image processing, and interpretable methods for students. Moreover, the project has the potential to increases research capacity and collaborations to generate new research opportunities for students from underrepresented communities to pursue advanced degrees in STEM.The project develops data-driven algorithms addressing interpretable cell quantitation problem. Facilitated by novel deep learning algorithms, the project harnesses the potential of big data to derive reliable, interpretable, and accurate evaluation. In addition, the project explores a generalized framework for extending existing findings and incorporating domain-expert knowledge to complement the modeling and learning process. Specifically, this research uses microscopy images, machine learning, transfer learning, and domain knowledge in three thrusts: (1) A deep learning based cell quantitation algorithm. The algorithm is featured with interpretability, guided by domain-expert knowledge for trust, reliable, and accurate performance. (2) Enhancing data-driven techniques in the area of transfer learning to fill important knowledge gaps at the intersection of the deep learning and biomedical imaging. (3) Advancing domain knowledge incorporation and human intervention methods for developing and operating deep learning based system for biomedical research and healthcare. All the research outcomes are made publicly available to facilitate extending and accelerating varied application development from diverse communities of researchers and students.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.