PRATIK SHAH

Research led by Dr. Shah and his laboratory has developed new paradigms for utilizing low-cost images captured through simple optical methods, making point-of-care clinical diagnosis more accessible and reducing reliance on specialized medical imaging devices, as well as complex biological and chemical processes. Dr. Shah’s lab has created interpretable systems and methods for the clinical translation of generative, predictive, and classification deep learning AI algorithms that extract diagnostic information from cells, tissues, and organs.

Example Projects and Collaborations:

• Advances in Uncertainty Quantification for Medical Image Segmentation: Two studies presented at the 2024 IEEE International Conference on Machine Learning and Applications (ICMLA), which focus on uncertainty quantification in deep learning models (DLMs) for skin cancer and prostate cancer image segmentation. Paper 1: “Uncertainty Quantified Deep Learning and Regression Analysis Framework for Image Segmentation of Skin Cancer Lesions” introduces a novel framework that integrates pixel-by-pixel uncertainty estimation into skin cancer lesion segmentation. This method helps clinicians identify regions requiring further review, increasing trust in AI models for clinical decision making. Paper 2 “Deep Learning with Uncertainty Quantification for Predicting the Segmentation Dice Coefficient of Prostate Cancer Biopsy Images” describes uncertainty quantification for prostate cancer biopsy image segmentation, predicting Dice coefficients and helping clinicians assess model performance at the image sub-region level. This study demonstrates how uncertainty maps can serve as a more reliable evaluation metric for clinical deployment.
• Workshop Paper and Special Session on Responsible Deep Learning for Software as a Medical Device: Dr. Shah conceptualized and led a special session at the IEEE International Symposium on Biomedical Imaging (ISBI 2022) focused on diversity, equity, and inclusion in medical imaging for uncertainty quantification and fair deep learning. This session featured experts from UCSF, the US FDA, and Dr. Shah, discussing the integration of unbiased imaging data, including race and ethnicity representation, into deep learning models to improve their fairness and clinical applicability. The workshop also highlighted strategies for developing high-performance, adaptable models that account for uncertainty and their regulatory pathways for real-world medical use. The extended paper, published on arXiv, explores these topics in-depth, detailing performance evaluations of AI/ML models in skin (RGB), digital pathology, and medical imaging (skin, kidneys, and prostate). The research aims to ensure safe and equitable deployment of AI-driven diagnostic tools (Paper on arXiv, IEEE Special session, Session Talk Video, Project Page)
• Improving Generalizability and Interpretation of Deep Learning Models: Deep learning models for segmenting complex patterns from medical images often suffer from poor generalizability. A study published in Cell Reports Methods presented an end-to-end toolkit that improves the generalizability and transparency of clinical-grade deep learning models. This toolkit helps researchers and clinicians identify hidden patterns in images and address issues related to the under-specification of clinical labels, reducing false positives and negatives in high-dimensional systems. While the study focused on medical images, the methods also apply to general image segmentation tasks, with wide applications in biomedical research and regulatory science (Project and publication link).
• Computational Haematoxylin and Eosin (H-&-E) Dye Staining and Destaining of Images of Tissue Biopsies: In collaboration with Brigham and Women’s Hospital, Dr. Shah’s team developed a “Computational H-&-E Dye Staining” system to digitally apply H-&-E dyes to photographs of unstained tissue biopsies for cancer diagnosis. The research also introduced a “Computational Destaining” algorithm, which removes dyes from stained tissue images, allowing reuse of biopsy samples. This innovative system uses neural networks to enhance the efficiency of diagnosing cancer while saving valuable patient samples (Project and publication link).
• Generative AI for Staining and Destaining of Images of Tissue Biopsies: A collaboration with Stanford University and Harvard Medical School led to breakthroughs in high-fidelity, explainable computational staining of tissue samples. The research focused on deep learning models that accurately “stain” non-stained tissue images to resemble computational H-&-E dye staining, and developed a novel loss function that preserves tissue structure. The models were shown to generalize well to unseen data, enabling accurate staining of patient samples and tumor grades not part of the training data. Additionally, neural activation maps were generated to enhance explainability, offering insights into the decision-making process of these models. This research was evaluated by pathologists for prostate cancer diagnosis and published in JAMA Network and covered by MIT News, demonstrating the clinical utility of virtual staining in decision-making (Project and publication link, MIT News story).
• Tumor Localization and Classification Using Computational H-&-E Dye Stained Images: Building upon the generative AI-based staining methods developed in previous work, Dr. Shah’s lab extended the use of computational H-&-E dye stained images for tumor localization and classification. In this project, non-stained biopsy images were first processed through generative models to create virtual H-&-E stained images, which were then classified using a ResNet-18 deep learning classifier to identify tumor regions. A gradient backpropagation (GBP) algorithm was employed to further localize tumor regions on the generated images, producing saliency maps that highlight areas of interest for clinicians. This research was selected for a Selected for a deep-dive spotlight session at the International Society for Optics and Photonics (SPIE) medical imaging conference for demonstrating how the computational H-&-E dye staining technique can be seamlessly integrated with tumor detection workflows, improving diagnostic accuracy, and extending the utility of virtual staining methods for real-time clinical applications (Publication link).
• Sepsis Diagnosis Using Dark Field Imaging: In collaboration with Beth Israel Deaconess Medical Center, Dr. Shah’s research explored the use of dark field imaging of capillary beds under the tongue to diagnose sepsis (bloodstream infection). A neural network developed by the lab achieved over 90% accuracy in distinguishing between septic and non-septic patients, providing a rapid method to assess sepsis and reduce unnecessary antibiotic use, helping to combat antimicrobial resistance (Project and publication link).
• Fluorescent Porphyrin Biomarkers for Cancer and Periodontal Disease: Dr. Shah led studies showing that fluorescent porphyrin biomarkers, which are typically used for imaging tumors and periodontal diseases, can be successfully predicted from standard white-light photographs. This method eliminates the need for specialized fluorescent imaging at the point-of-care, making diagnosis more affordable and accessible (Project and publication link).
• Automated Segmentation of Oral Diseases: In another study, Dr. Shah’s lab trained neural networks to automatically segment oral diseases from standard white-light photographs. These automated systems also correlate disease patterns with systemic health conditions, such as optic nerve abnormalities, to generate personalized risk scores for patients.(Project and publication link)

These examples show Dr. Shah’s lab’s contributions to the next generation of computational medicine, particularly in the development of generative AI models for medical imaging. Our goal is to assist physicians and patients in making more accurate, efficient, safe and cost-effective diagnoses at the point of care..