Eric Lefkofsky’s Post

Early Access Paper Just Posted: An interpretable medical visual question answering model that uses the counterfactual causal-effect intervention strategy (CCIS-MVQA). The CCIS-MVQA introduces the layer-wise relevance propagation to automatically generate counterfactual samples and applies counterfactual causal reasoning to enhance interpretability and generalization of the model. Read the paper: https://lnkd.in/gmWF9vuW

PhysChem property prediction modules in Percepta offer: 🏆 Deeper insights 🏆 Trainability with experimental data 🏆 Expandability to third-party models — all in a single, easy-to-use suite. Learn more ⬇️

"How can you support our multi-reader trials/research?" Was one of the top questions asked on the booth at #DDW2024 👉 The answer is: on Entrolytics with structured reporting! 👩🏼💻 Entrolytics: Motilent's cloud-based, multi-modality central reading platform built for everything GI ✅ Industry standard and configurable #eCRF eg. MaRIA, IBUS-SAS, Magnifi-CD ✅ Annotation tools auto populate eCRF fields > fewer clicks and no rekey errors ✅ #Multireader / adjudication support ✅ Immutable published reports: Once published reports are secure Check out the sMARIA score within our structured reporting workflow below 👇 #IBD

We want the machines to understand what we need and ready to hgive us the output in a domain so that we could find a lot of insights in the medical research It should be a. Twoway communication building trust in the system Verify for deciding Diagnostic or forensic analysis Counter fact, reason New medical insights

https://lnkd.in/g5nzFtKq We developed a volumetric quantitative evaluation software called Vector Volume Histogram (VVH) to evaluate respiratory induced organ motion using deformable image registration!!

Digital PCR, a meticulously crafted technique, unveils a captivating realm where each cluster holds unique significance and interpretation. Embark on a journey into this fascinating scientific domain and discover the profound insights it offers. https://lnkd.in/gfcSWZFg #DigitalPCR #DQ24

Spotlight at NeurIPS 🌟 🎉 Giving classifiers the option to reject cases is crucial for reliable application. Yet, current evaluation in this field, including primary metrics like AURC, is flawed. We systematically anlayze and overcome these pitfalls: https://lnkd.in/d6DPwTpK work led by Jeremias Traub, with Carsten Lüth, Till Bungert, Lena Maier-Hein, Klaus Maier-Hein, Michael Baumgartner.

EASY UNDERSTANDING OF LABORATORY DATASET SDTM #CLINICALSASPROGRAMMER #SASPROGRAMMER

🚨New Pre-print out! MemControl: Mitigating Memorization in Medical Diffusion Models via Automated Parameter Selection. A new strategy to mitigate memorization in Diffusion models We target the problem of memorization in diffusion models and introduce a new framework that identifies the optimal set of parameters to be fine-tuned for two objectives: 1. Improving image generation performance 2. Reducing memorization Our results show that fine-tuning just 0.019% of the model parameters can enable high-quality image generation with minimal memorization! Arxiv Link: https://lnkd.in/giWr2_2X Here is a Twitter thread explaining further details and results: https://lnkd.in/gUaxTk3j Work done jointly with Pedro P. Sanchez, Ondrej Bohdal, Sotirios (Sotos) Tsaftaris and Timothy Hospedales

Our New Model and Latest Paper: "Advanced ML Approaches to PGx Recommendations in Precision Medicine" This summer at the 2024 IEEE 48th Annual Computers, Software, and Applications Conference (IEEE), we, along with Nisha Talagala and Danika Gupta, introduced PGxAI Sirius, our new model designed to predict pharmacogenetic (PGx) guidelines with enhanced accuracy. Our model leverages a novel dataset from the PharmGKB knowledge base, incorporating chemical structure, genetic information, and recommendation annotations. After evaluating several models—including Multi-Layer Perceptron, K Nearest Neighbors, Random Forest, Logistic Regression, Linear SVC, and XGBoost—XGBoost emerged as the top performer, combining predictive power with explainability. It achieved an accuracy of 89.14%, with F1 scores from 0.85 to 0.90, and precision and recall ranging from 0.83-0.97 and 0.82-0.97, respectively. For the full text of our extended abstract and model description, read more on the link. And stay tuned for our next model, PGxAI Vega, with 10B parameters, launching at the beginning of December. To follow our progress and support us, hit the fire button below! 😎