Current Challenges and Opportunities in Implementing AI/ML in Cancer Imaging: Integration, Development, and Adoption Perspectives

Abstract

AI and ML can process imaging data quickly and with a high degree of accuracy, potentially surpassing what humans can achieve. Nonetheless, there are challenges in applying these algorithms to cancer imaging.  An analysis of both the constraints and the potential advancements in this field offers an informed perspective on the present status of technological and research developments in AI and ML applied to cancer imaging. This study presents an in-depth analysis of the challenges and opportunities associated with the AI/ML in cancer imaging, focusing on three critical areas: integration, development, and adoption.  In the integration phase, the study addresses the issues in managing and harmonizing the influx of diverse biomedical data, including multi-modal imaging, multi-omics, and electronic health records. The paper emphasizes the importance of such integration for personalized medicine and precision oncology, in cancer image analysis and the understanding of cancer biology for treatment responses. However, challenges such as data quality, diversity, and the need for robust computational methods like transfer learning and domain adaptation to ensure generalizability across studies are highlighted. The development phase discusses the necessity of collaboration from distinct disciplines, particularly the involvement of clinicians AI tools development, to ensure that they are clinically relevant and fit seamlessly into existing healthcare systems. Challenges in developing reproducible AI algorithms for tumor segmentation, diagnosis, and the identification of biomarkers are examined. The study also explores the implications of deep learning success despite data annotation challenges, advocating for a shift towards models that require minimal supervision. The need for AI education within the radiological workforce and the role of informatics teams in AI tool development and testing are also discussed. In the the adoption phase, the paper discusses the growing demand for imaging services amidst workforce shortages, emphasizing the need for AI/ML solutions to alleviate radiologist stress and burnout. It critically examines radiologists' perceptions of AI and ML, including the challenges posed by the black-box nature of AI models. The paper advocates for the development of explainable AI to enhance patient safety, model robustness, and end-user trust, while also underscoring the importance of educational investments, tool testing, data curation, and vendor collaboration.