Paris – Nov 13, 2019 – The application of Convolutional Neural Networks (CNN), an advanced form of Artificial Intelligence (AI), has been widely used in research for automating cytology images. Most of the research to date however, focuses on training the Neutral Networks to classify individual cells or regions of the image. Recent advances in […]lire la suite
Paris – Dec 12, 2016 – Keen Eye will finalize the development of its innovative image interpretation platform for laboratories and biomedical research Keen Eye, a company that designs, develops and markets innovative image analysis solutions for research and the medical sector, announced today a 1.5 million fund raising from Seventure Partners’ Quadrivium seed fun. […]lire la suite
Paris – April 6, 2019 – KEEN EYE laureate of the 7th edition of the Concours d’Innovation Numérique – CIN – « Live Better » category The solution developed by Keen Eye, a French startup specialized in digital technology, has caught the attention of the jury* of the French 7th edition of the Concours d’Innovation […]lire la suite
JP Morgan / Biotech Showcase San Francisco, Jan 12-15, 2020 – Sylvain Berlemont, CEO of KEEN EYE, will attend JP Morgan. Feel free to send us an invite on the Partnering platform. image : https://drive.google.com/open?id=12cimYN1Qxnzsz-FWEQ-CxZj2LAB8K7W4 EVENTS #2 Biomarker Series Manchester, Feb 18-20, 2020 – Let’s meet us at Biomarker Series in Manchester to discuss how […]lire la suite
PARIS – Keen Eye, a French technology company specialized in image analysis for the life science industry, announced today a strategic partnership agreement with Iris Pharma, a world-wide leading ophthalmology-focused contract research organization (CRO) that offers preclinical and clinical drug / device development services. The partnership aims at providing Iris Pharma artificial intelligence (AI) applications […]lire la suite
Histological slides of mice with inflammatory bowel syndrome were analyzed by our AI model to score the severity based on cellular infiltration. Data scientists at Keen Eye used the multiple instance learning (MIL) method to divide up a large slide into small tiles for training the AI. The self-learning algorithm scores each tile for its contribution to the final classification. A heatmap is then generated to reveal its decision-making process, making this an excellent tool for biomarker discovery projects. Read more here.
We developed a deep learning training strategy to improve the classification of squamous cells for cervical cancer screening. Traditional deep learning models have huge drawbacks in the misclassification between normal and carcinoma-in-situ, the most severe form of the cancer. We evaluated a new architecture that takes into account the distances between classes and introduced penalties depending on the severity of the mistakes. The model showed a 94% accuracy on normal/abnormal classification. Furthermore, to explain the model’s prediction, we computed each pixels contribution to the prediction. Strong contribution from the nucleus region confirms that the model has learned the relevant features that match with medical consensus. Read more here.
Developing computer algorithms that use classic feature analysis to classify among lung cancer subtypes of adenocarcinoma and squamous carcinoma, requires that pathologists annotate hundreds of cells on the training slides. Here we demonstrate the effectiveness of training the deep learning AI model on the whole slide image: using only one annotation – the slide-level diagnosis. Furthermore, we demonstrate that AI can accurately highlight the regions of interest (hotspots) where the pathologist can investigate further for insights. Read more here.
Automating the analysis of IVCM (in vitro confocal microscopy) images can aid in the early detection of Sjögren syndrome, and prevent complications. The technical challenge, however, lies in the limited available data for training the deep learning model. Keen Eye researchers collaborated with clinicians at the Quinze-Vingts National Ophthalmology Hospital in Paris to design an innovative hybrid transfer learning model. By directing the AI model to focus on learning relevant features, we improved the model’s ability to converge using a small training sample from 80 patients. The resulting predictive accuracy of 81% shows great promise with this technique. Read more here.