Accepted Posters & AI Demos
Breaking the Barriers to Effective Type 2 Diabetes Care: An NLP-Enabled Decision Support System Solution
Valentina Estupiñán-Vargas, Andrés Montoya-López
A decision support system for type 2 diabetes care, using NLP on unstructured data and clinical expertise. Provides personalized recommendations based on real-time analysis of patient data from health records. Aims to enhance treatment quality and efficiency, reduce costs, and bridge the gap in specialized healthcare personnel.
Characterizing tissue composition through combined analysis of morphologies and transcriptional states
Feng Bao, Steven Altschuler, Lani Wu
Advances in spatial transcriptomics technologies enable simultaneous profiling of morphological and transcriptional modalities from the same cells or regions within tissues. We present multi-modal structured embedding, an approach to deeply characterize tissue heterogeneity through analysis of combined image and transcriptional measurements.
Diagnostic Accuracy of Synthetic FDG PET Images from MR Imaging
Helena Zhang, Jiahong Ouyang, Jarrett Rosenberg, Greg Zaharchuk
Using a deep learning model, FDG PET images have been synthesized from from multi-contrast MRI. This study aims to evaluate the diagnostic accuracy of synthetically generated FDG PET images compared to actual FDG PET images, in predicting cancer recurrence, using the biopsy-proven and clinically proven diagnoses from MRI images as the reference standard.
Do Good with Data: Unlocking the potential of AI for Nonprofits
Do Good with Data, Inc
Data Science and Artificial intelligence (AI) can be an effective and powerful tool for non-profit organizations. With latest AI technology can be used by nonprofits to effectively increase their outreach and further achieve their goal than what was believed before.
Physics Informed Deep Residual Networks Based Type-A Aortic Dissection Prediction
Joy Cao, Dr. Min Zhou
Acute Type-A aortic dissection is a life-threatening condition. The proposed novel physics-informed deep residual network shows great potential in creating a cost-effective non-invasive predictor framework. We believe that by deploying this predictor, doctors can take appropriate early actions and greatly reduce the mortality of Type-A aortic dissection patients.
Predicting FDG-PET Images from Multi-contrast MRI using Deep Learning in Patients with Brain Neoplasms
Jiahong Ouyang, Kevin T. Chen, Rui Duarte Armindo, Guido Alejandro Davidzon, Elizabeth Hawk, Farshad Moradi, Jarrett Rosenberg, Ella Lan, Helena Zhang, Greg Zaharchuk
FDG PET is valuable for determining presence of viable tumor, but is limited by geographical restrictions, radiation exposure, and high cost. On the other hand, MR is widely available and non-invasive. Thus, in this project we aim to generate diagnostic-quality PET equivalent imaging for patients with brain neoplasms by deep learning with multi-contrast MRI.
Predicting Novel Drug-Drug Interaction Risks with a Machine-Learning Framework: a Case Study in SARS-CoV-2 Antivirals
Dean Wang, Danielle Wang
To improve safety and efficacy in SARS-CoV-2 treatment, we developed a pharmacologic profile-based ML algorithm to predict undiscovered drug-drug interactions for three antivirals. All drug combinations were categorized by severity and validated with FAERS. With high predictive accuracy, this model may streamline prescriptions and accelerate novel drug development.
Puncture point identification in PCNL surgery using epipolar geometry
Pinak Paliwal, Palani Narayan
There exists a problem of determining the 3D orientation of puncture needle to renal collecting system in Percutaneous nephrolithotomy (PCNL) surgery. This can be modeled through epipolar geometry to pick the correct orientation. It is possible to determine a good orientation based on camera geometry and training data based on human doctors’ choices.
Simple Hardware-Efficient Long Convolutions for Sequence Modeling
Daniel Y. Fu, Elliot L. Epstein, Eric Nguyen, Armin W. Thomas, Michael Zhang, Tri Dao, Atri Rudra, Christopher Ré
What is the simplest architecture you can use to get good performance on sequence modeling? We study whether directly learning long convolutions over the sequence can match the performance of State Space Models (SSMs) and Transformers, while maintaining a sub-quadratic compute scaling in the sequence length.
Almanac: A Knowledge Grounded Large Language Model for Physicians
William Hiesinger, Cyril Zakka, Akash Chaurasia, Rohan Shad, Alex Dalal, Jennifer Kim, Michael Moor, Kevin Alexander, Euan Ashley, Jack Boyd, Kathleen Boyd, Karen Hirsch, Curtis Langlotz, Joanna Nelson
Despite many promising applications in clinical medicine, adoption of Large Language Models (LLMs) in real-world settings has been largely limited by their tendency to generate incorrect and sometimes even toxic statements. In this study, we develop Almanac, an LLM framework augmented with retrieval capabilities for medical guideline and treatment recommendations.
BREAST AI: A Low Cost, Explainable Artificial Intelligence Based App for Efficient Diagnosis of Breast Cancer in Developing Areas
Accurate diagnosis of breast cancer is critical to successful treatment; however, those living in developing settings often don’t have access to resources such as mammograms and doctors to get a precise diagnosis. BREAST AI has the potential to diagnose breast cancer in developing settings based on ultrasound scans in a fast, explainable, accessible, and accurate way.
Dimble: Orders of Magnitude Faster Medical Imaging IO
Chris Ayling, Ben Sand, Tanisha Banaszczyk, Akshay Chaudhari, Rogier van der Sluijs, Arjun Desai, Sarthak Pati, Jake Goulding
Dimble is an open source fast DICOM/NIfTI replacement. It is a new format, with easy conversion to/from DICOM. This makes working with medical data for machine learning at scale more efficient and much simpler.
DrAid Liver Cancer CT- Detecting early signs of Liver Cancer
Steven Truong, VinBrain
DrAid- Liver Cancer CT: An AI platform enabling segmentation & early cancer detection of all types of tumors and classifying them into 4 four different classes: HCC, other malignant (other than HCC), benign, and ambiguous., thereby assisting clinicians to speed up medical management.
Isabel Hyo Jung Song, Aarav Sharma, Thuy-Anh Nguyen
Saccador effectively leverages computer vision algorithms such as BiseNet and MTCNN to establish a correlation between eye-gaze estimation and saccadic behavior. Saccador offers a novel and unique approach to correlating eye movement to the possibility of abnormal saccades.
UMLS Entity Link/Extraction from Unstructured Text (Over 400k concepts)
Santosh Gupta, Science.IO
Science.IO's premier model, Kepler, can map medical terms from unstructured text to over 400,000 entities in just one API call. Kepler identifies all entity spans within the input text, then links them to established medical ontologies. At present, Kepler offers support for the following medical ontologies: UMLS, ChEMBL/ChEBI, dbSNP, Cell Line Ontology (CLO), GeneID, ClinVar, and NCBI Taxonomy ID