15th International Conference on
Focused Topic: Deep Learning in Biology and Medicine
Deep Learning for Imaging Diagnostics
The field of analyzing medical images with computers started in the early 1960s and for half a century the goal of achieving automated analysis at the level of human experts (radiologists, pathologists, ophthalmologists, etc) remained elusive. This changed in 2013 when it became clear that deep learning is a technology that outperforms all traditional techniques in image processing, including the dominant machine learning approach of had crafted feature extraction followed by statistical classification by, e.g., support vector machines, random forests etc. In the first part of my lecture, I review some of the traditional approaches and provide examples of recent deep learning based solutions that perform on par with human experts, or even outperform these human experts for various clinically relevant medical image analysis tasks. In the second part of my lecture, I discuss the implications of these exciting developments for diagnostic medicine in the next few decades.
Download PDF [16 MB]
Focused Topic: Knowledge Discovery in Human Behavior Recognition
Learning to understand Human Behaviour: in multimedia, surveillance and automotive
Università di Modena e Reggio Emilia
Human behavior understanding (HBU) is a central topic in computer vision and pattern recognition. New Deep learning paradigms, real and synthetic available datasets and GPUs facilities improved strength, robustness and accuracy substantially.
HBU research produced a seamless ocean of projects and results, covering aspects of people detection and body part detection (face, hands..), of tracking, of action, expression, emotion and activity recognition. The talk will focus on some HBU aspects only, related with the detection of humans and their activity under different contexts, carried out at Aimage Lab of university of Modena and Reggio Emilia, italy recently. In multimedia HBU has been studied in unknown and unconstrained video footage for action analysis and recently for video action description: we will discuss on novel applications of video captioning with naming in edited video. Similarly in surveillance, more than ten years of research achieved impressive results in people detection and tracking, as a starting point for analysis of action and behaviour from fixed camera. Now new impressive results are carrying out with large datasets and Deep Learning. We will discuss on some new architectures of for people pose understanding in the wild, covering occlusions and crowd . Finally, we will consider the automotive field, a key application area nowadays, due to the fact that cameras are considered mandatory sensing components in cars to support assisted or automatic guidance, to improve the safety and comfort of drivers and passengers. HBU is needed inside and outside the car: to understand what the driver is doing, what he/she can do to drive better and what people around the car are doing. Here we will present some results achieved in collaboration within Ferrari RedVision lab for Human-Vehicle- Interaction applications. These solutions are mostly based on 3D information, saliency analysis and video segmentation and ground on extensive use of deep networks technology. I will discuss deep architectures we propose for understanding person presence, person behaviour and person interaction.
Focused Topic: Retinal Image Analysis for Screening and Diagnosis
Eyes are the windows of the body: the analysis of corneal and retinal images
BioImLab, Department of Information Engineering
Unviersity of Padova
A paraphrase of the well-known metaphor “Eyes are the windows of the soul” can effectively highlight the clinical importance of analyzing images from the eye, namely corneal and retinal images. In this talk the various algorithms for the automated analysis of these images, developed over the years at BioImLab, will be presented. Corneal images from both specular and confocal microscopy have been addressed, whereas the retinal images have been acquired with conventional or pediatric fundus cameras. Techniques from classical image processing and analysis, machine learning and pattern recognition, statistical analysis, and others have been successfully applied. The aim of these analyses was to provide ophthalmologists with a quantitative description of the main clinical parameters used in their diagnostic procedures. Clinical reliability, total run-time, and user-friendliness have thus been the key features taken into account during the development. Tools to analyze several of the cornea layers (epithelium, sub-basal nerves, stroma, endothelium) and different features of retinal images will be described and their performance assessed by comparing the results, wherever possible, with manual analysis. Some hints about work currently in progress will also be provided to give a perspective of possible future developments.
Download PDF [10 MB]
ICIAR - International Conference on Image Analysis and Recognition
Send your comments to firstname.lastname@example.org