Project Grant F31AG094285

DEEP LEARNING FOR HIGH-THROUGHPUT HISTOPATHOLOGY MAPPING IN ALZHEIMER'S DISEASE - PROJECT SUMMARY/ABSTRACT ALZHEIMER'S DISEASE (AD) IS A NEURODEGENERATIVE DISEASE ASSOCIATED WITH AGING, DISTINGUISHED BY THE PRESENCE OF AMYLOID PLAQUES AND NEUROFIBRILLARY TANGLES (NFTS). THE ECONOMIC AND SOCIAL BURDEN ASSOCIATED WITH AD IS PROJECTED TO INCREASE SUBSTANTIALLY OVER THE NEXT SEVERAL DECADES. WHILE SEVERAL MONOCLONAL ANTIBODY MEDICATIONS FOR AMYLOID PLAQUE CLEARANCE HAVE BEEN APPROVED AND SHOW AN EFFECT ON COGNITIVE DECLINE, THEIR EFFECTIVENESS IS INCOMPLETE AND VARIES AMONG THE DIFFERENT DRUG TREATMENTS. ONE POSSIBILITY IS THAT EACH DRUG PREFERENTIALLY TARGETS DIFFERENT AMYLOID PLAQUE SUBTYPES THAT CONTRIBUTE TO COGNITIVE IMPAIRMENT. A BETTER CHARACTERIZATION OF AD PATHOLOGY SUBTYPES AND THEIR DISTRIBUTION WITHIN THE BRAIN IS CRITICAL FOR DEVELOPING MORE EFFECTIVE TREATMENTS. WHILE PATHOLOGY IS TRADITIONALLY PERFORMED BY A TRAINED PATHOLOGIST AT A MICROSCOPE, ADVANCES IN TECHNOLOGY NOW ALLOW FOR SCANNING AND STORAGE OF WHOLE SLIDE IMAGES (WSIS) AND COMPUTATIONAL QUANTITATIVE ANALYSIS. IN OTHER FIELDS OF HISTOPATHOLOGY, COMPUTATIONAL METHODS ARE USED FOR CLASSIFICATION OF CANCER TISSUE AND SEGMENTATION OF CELLS. SOME EFFORTS HAVE BEEN MADE TO USE QUANTITATIVE METHODS ON AD WSIS, BUT APPLICATION OF COMPUTER VISION TECHNIQUES TO NEUROPATHOLOGY LAGS BEHIND PROGRESS IN OTHER MEDICAL IMAGE FIELDS. HIGH THROUGHPUT, QUANTITATIVE METHODS FOR HISTOPATHOLOGY ANALYSIS IN AD WILL HELP CHARACTERIZE AD SUBTYPES AND REGIONAL SUSCEPTIBILITY TO PROTEIN AGGREGATION. AS PART OF THIS F31 TRAINING GRANT, I PROPOSE TO DEVELOP NEW TRAINING IN DEEP LEARNING AND ARTIFICIAL INTELLIGENCE TECHNIQUES FOR CHARACTERIZATION OF AD PATHOLOGY. PRELIMINARY WORK IN OUR LAB USED MANUAL SEGMENTATION OF BRAIN SUBREGIONS AND MACHINE LEARNING-ASSISTED ANNOTATION OF PATHOLOGY TO QUANTIFY REGIONAL DISTRIBUTION OF AMYLOID AND TAU. WE USED THIS WORKFLOW TO FIND DIFFERENCES IN DISTRIBUTION OF HIPPOCAMPAL TAU BETWEEN TWO TAU STAINS AND CHARACTERIZES TAU IN A UNIQUE FAMILIAL AD SUBPOPULATION. DEVELOPING MORE AUTOMATED COMPUTER VISION APPROACHES WOULD INCREASE EFFICIENCY AND ACCURACY OF SEGMENTATION. FIRST, IN AIM 1 WE WILL USE DEEP LEARNING (DL) FOR SEGMENTATION OF DIFFERENT PATHOLOGY SUBTYPES IN WSIS, SUCH AS DENSE VERSUS DIFFUSE AMYLOID PLAQUES. THIS WILL USE A U-NET TRAINED ON THE CONSENSUS OF THREE SEPARATE ANNOTATIONS. THE OUTPUT OF QUANTIFICATION CAN BE USED TO ASSESS RELATIVE DENSITY OF TAU AND AMYLOID SUBTYPES ACROSS FIVE REGIONS OF THE CORTEX. AIM 2 WILL USE A COMBINATION OF AN AD CLASSIFICATION MODEL TRAINED ON HIPPOCAMPAL AND PREFRONTAL CORTEX SECTIONS AND EXPLAINABLE ARTIFICIAL INTELLIGENCE (XAI) TO PROVIDE A HEATMAP OF SUBREGIONS THAT ARE IMPORTANT FOR DISTINGUISHING NORMAL COGNITION FROM AD. THESE EXPERIMENTS COLLECTIVELY WILL PROVIDE ME WITH CRITICAL CAREER TRAINING AND DEVELOP NOVEL METHODS FOR NEUROPATHOLOGY ANALYSIS TO PROVIDE CELLULAR-LEVEL DATA ON DISTRIBUTION OF PATHOLOGY IN AD AND ANATOMICAL FEATURES OF DISEASE PROGRESSION.