Multimodal Biomedical Image Classification and Retrieval with Multi Response Linear Regression (MLR)-Based Meta Learning

This paper presents a classification-driven biomedical image retrieval approach by combining multiple visual and text features with a multi-response linear regression (MLR)-based meta-learner. Feature descriptors at different levels of image representation are often in diverse forms and complementary in nature. For modality detection of medical images, the MLR has been proposed as a trainable combiner for fusing class probability outputs of several base-level SVM classifiers on different visual and text features as inputs. The advantage of using MLR here over other generalizers is its interpretability as the weights generated by it indicate the different contributions that each features makes for class prediction. Hence, a query-specific adaptive similarity fusion approach is also proposed for image retrieval. Based on the on-line prediction of the query image modalities, individual feature weights generated by MLR are used in a linear combination of similarity matching function for image retrieval. The classification and retrieval results were evaluated evaluated on a standard ImageCLEFmed’2010 benchmark data set of 77,000 images with associated XML annotations and it showed improved performances.