DeepSRH: Transforming Molecular Cancer Screening with Stimulated Raman Histology and Deep Neural Networks

About This Grant

ABSTRACT Molecular markers have transformed the diagnosis, prognosis, and treatment of human cancers. However, ac- cess to molecular testing is uneven and often delayed because of the complex laboratory infrastructure required to obtain molecular data from surgical cancer specimens. These and other multifaceted barriers result in <10% of patients receiving molecularly targeted therapies that improve patient survival. Rapid, point-of-care molecular screening would transform the treatment of human cancers by identifying patients at the earliest possible point in their cancer care, ensuring easy access to molecular testing and allowing for immediate delivery of molecu- larly targeted surgical and medical treatment. There exists a critical need for innovative and scalable methods for molecular cancer screening. Stimulated Raman Histology (SRH) is an emerging and innovative optical imag- ing method that produces fast, high-resolution, label-free microscopy images of fresh surgical specimens at the patient’s bedside. We recently demonstrated that SRH combined with artificial intelligence (AI) models can accu- rately detect tumor infiltration and predict the key molecular markers in brain tumors within two minutes of tumor biopsy (Nature Medicine 2023, Nature 2024). However, earlier SRH-AI models are limited in that they require extensive data annotations, lack clinical context, and cannot adapt to other organs or cancer types. Address- ing these limitations through advanced AI methods is an essential step towards developing scalable SRH-based molecular cancer screening methods. Here, we aim to develop DeepSRH, an integrated, point-of-care, SRH- based screening system for rapid and accurate molecular marker prediction using deep neural networks. Driven by our preliminary data and motivated by recent work on vision-language AI models, we hypothesize that large- scale, self-supervised foundation model training on a diverse SRH dataset plus efficient model fine-tuning can produce SRH-AI models for accurate molecular marker prediction. The central objective of this proposal is to vali- date DeepSRH across cancer types (brain, lung, prostate) for AI-based molecular cancer screening and real-time clinical decision support. Firstly (Aim 1), we aim will develop a self-supervised learning strategy, called Slide Pre- trained Transformers (SPT), for DeepSRH foundation model training. Next (Aim 2), we will optimize DeepSRH foundation models for molecular marker prediction through efficient and multimodal fine-tuning methods. Lastly (Aim 3), we will test fine-tuned DeepSRH performance in a multi-center, multi-organ cohort of cancer patients. Using College of Pathology and DECIDE-AI guidelines, our milestone is a balanced diagnostic accuracy of ≥95% with a calculated target sample size of 424 total patients. The expected contribution is an autonomous SRH-based molecular cancer screening workflow. This proposal aims to bridge the gap between rapid optical imaging and real-time molecular screening by strategically addressing multiple challenges using state-of-the-art AI techniques. Translating SRH using advanced AI could revolutionize access to molecular testing in today’s precision medicine landscape. We aim to create a new standard for the accessibility of molecular diagnosis in human cancers.