What I Found in AI – Mar. 18th to March25th
Every week, I gather and summarize the most interesting AI-related articles I come across from various sources.
In this week’s edition of the AI newsletter, we’re diving into some seriously fascinating developments across the AI landscape. DeepSeek’s V3 model now runs more efficiently on consumer hardware and is open for commercial use under the MIT license. Over in healthcare, a breakthrough ECgMLP model is diagnosing endometrial cancer with 99.26% accuracy—an astounding leap from current methods. Microsoft’s new Data Formulator tool makes data visualization easier and more powerful than ever with a no-code, AI-enhanced interface. A South Korean study highlights how generative AI is improving chest x-ray reading times and diagnostic accuracy, while researchers at Stanford and Cornell introduce MorphLDM, a model that’s reshaping how we generate 3D brain MRIs. We also tackle the thorny issue of medical hallucinations in AI, with a deep dive into their risks, detection strategies, and ethical implications. Let’s get into it!
DeepSeek releases improved V3 model under MIT license
Summary
DeepSeek has released an improved version of their DeepSeek-V3 language model under the MIT license, featuring better performance and efficiency. The model can run on high-end consumer hardware and shows improved programming capabilities while requiring less computational resources than traditional LLMs.
Main Points
DeepSeek-V3 switched from custom license to MIT License, allowing more flexible commercial use
The new version shows improved programming capabilities with a 60% benchmark score
Model uses 671B parameters but only activates 37B when processing, improving efficiency
Can run on consumer hardware (Mac Studio) with four-bit quantization at 20 tokens per second
Was trained on 14.8 trillion tokens using 2.8 million GPU hours and fine-tuned with DeepSeek-R1
ECgMLP: A novel gated MLP model for enhanced endometrial cancer diagnosis
Summary
Researchers have developed an AI model called ECgMLP that can detect endometrial cancer with 99.26% accuracy by analyzing histopathological images. This represents a significant improvement over current automated diagnosis methods, which have accuracy rates between 78.91% and 80.93%. The model has also shown promising results in detecting other types of cancer.
Main Points
The AI model achieves 99.26% accuracy in detecting endometrial cancer, significantly higher than existing automated methods
The model works by enhancing and analyzing histopathological tissue images
The technology can be applied to other types of cancer detection, including colorectal (98.57%), breast (98.20%), and oral cancer (97.34%)
The system could be integrated into clinical practice to assist doctors with cancer diagnosis decision-making
Endometrial cancer is the most common gynecological cancer in Australia
Microsoft released a new AI powered data analysis tool
Summary
Microsoft has released Data Formulator, an open-source AI-powered data analysis tool that combines UI interactions with natural language inputs to create advanced visualizations. The tool allows for data transformation and computation without coding requirements.
Main Points
Data Formulator is a no-code, AI-powered tool for creating rich data visualizations
The tool combines traditional UI interactions with natural language processing for better user experience
Key features include beyond-dataset computations, intuitive field selection, AI-powered insights, and interactive editing
The tool is completely open-source and available on Github
It offers real-time chart refinement and visualization capabilities through follow-up prompts
Generative AI model validated in chest x-ray study
Summary
A South Korean study validates a generative AI model that improves chest x-ray reporting efficiency and accuracy. The study showed reduced reading times by 14 seconds per image and increased detection sensitivity for certain abnormalities when radiologists used AI-generated preliminary reports.
Main Points
Average reading time reduced from 34.2 to 19.8 seconds with AI assistance
Improved sensitivity in detecting specific abnormalities like widened mediastinal silhouettes and pleural lesions
Report agreement and quality scores improved with AI assistance
Current limitations include inability to compare with prior radiographs and consider clinical context
The technology shows promise for transforming radiologist efficiency and accuracy in the future
Generating Novel Brain Morphology by Deforming Learned Templates
Summary
Stanford and Cornell researchers have developed MorphLDM, a new AI model that generates highly accurate 3D brain MRIs by learning and deforming anatomical templates. The model outperforms existing methods like GANs and standard diffusion models, showing significant improvements in fidelity and morphometric accuracy.
Main Points
MorphLDM generates high-fidelity 3D brain MRIs using a novel deformable template approach
The model showed 22.6% improvement in FID scores and 15% better voxel-based morphometry accuracy compared to existing methods
Achieved 57% reduction in mean absolute error for age prediction accuracy
Uses a unique architecture combining latent diffusion models with deformation field processing
Addresses the challenge of generating synthetic medical data for clinical AI applications
Medical Hallucinations in Foundation Models and Their Impact on Healthcare
Summary
This research paper addresses the critical issue of medical hallucinations in AI foundation models and their impact on healthcare. The study provides a comprehensive analysis of how AI models can generate misleading medical information, potentially affecting clinical decisions and patient safety. The research includes a taxonomy of medical hallucinations, model benchmarking, and a multi-national clinician survey.
Main Points
Development of a taxonomy for understanding and addressing medical hallucinations in AI models
Benchmarking of models using medical hallucination datasets and physician-annotated LLM responses
Implementation of Chain-of-Thought and Search Augmented Generation techniques to reduce hallucination rates
Need for robust detection and mitigation strategies for medical hallucinations
Call for clearer ethical and regulatory guidelines to ensure patient safety in AI-integrated healthcare