Image Representation and Retrieval Using Support Vector Machine and Fuzzy C-means Clustering Based Semantical Spaces
(Professor, Concordia University, Canada)
The texture moments found from the medical image on the left are one of many features used for image query and retrieval.
LORIA-INPL-Université de Nancy, France
Professor Bhattacharya spoke about image representation and retrieval applied to the medical domain. Medical images can be collected from different technologies like X-ray, ultrasound, or magnetic resonance. These images contain information that can be crucial in building a diagnostic.
Text based searches of annotated images have many limitations. It is difficult to capture the rich content of an image using text. Moreover, different experts can see different things, and even the same expert can see different things at different times.
This implies a need for content-based image retrieval (CBIR). This involves different domains like machine learning, pattern recognition, computer vision or artificial intelligence. CBIR systems are usually built in three steps: feature extraction, representation, and matching.
There are limitations that we have to be aware of. First there is a sensory gap between the object in the world and the information in the description. Secondly, there is a semantic gap, that is disagreements between the information we can extract from the visual data and what the expert user expects to extract. In medical images, moreover, there are important features that are more local than global, and small variations in these may imply radically different diagnoses.
Bhattacharya proposed a CBIR system based on a mixed approach (text and image) called ImageCLEF, whose objective is to perform image retrieval and automatic image annotation. The system is made of different steps. First, there is a categorization of the images by medical domain in order to improve the performance. At this step, it is possible to make a search for an image modality (X-ray, ultrasound, ...), for an anatomic region (foot, heart, ...), or for a pathology (leukemia, myelogenous, ...). A learning-based approach is then used on these data by using a support vector machine. Then he used an unsupervised fuzzy c-means clustering on the pixel images using a combined feature vector.
The results are promising and additional research on how real users query image retrieval systems will shed light on which system-oriented evaluation measures are most important.