With the ever-increasing amount of annotated medical data, large-scale, data-driven methods provide the promise of bridging the semantic gap between images and diagnoses. The goal of this research project is to increase the scale at which interactive systems can be effective for knowledge discovery in potentially massive databases of medical images. Particularly, we focus on the automatic analysis of histopathological images, and propose a scalable image retrieval framework with high-dimensional features extracted in cell-level. We present a kernelized and supervised hashing method to bridge the semantic gap. With a small amount of supervised information, our method can compress a 10,000-dimensional image feature vector into only tens of binary bits with informative signatures preserved, and these binary codes are then indexed into a hash table that enables real-time retrieval. We validate the hashing-based image retrieval framework on several thousands of images of breast microscopic tissues for both image classification (i.e., benign vs. actionable categorization) and retrieval. Our framework achieves high search accuracy and promising computational efficiency, comparing favorably with other commonly used methods.
@ARTICLE{Zhang_TMI_hashing, author={Zhang, X. and Liu, W. and Dundar, M. and Badve, S. and Zhang, S.}, journal={IEEE Transactions on Medical Imaging}, title={Towards Large-Scale Histopathological Image Analysis: Hashing-Based Image Retrieval}, year={2015}, month={Feb}, volume={34}, number={2}, pages={496-506}, } @InProceedings{Zhang_2015_CVPR, author = {Zhang, Xiaofan and Su, Hai and Yang, Lin and Zhang, Shaoting}, title = {Fine-Grained Histopathological Image Analysis via Robust Segmentation and Large-Scale Retrieval}, booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, month = {June}, year = {2015} }