PhD Defense by Sailesh Conjeti

Speaker: Sailesh Conjeti Date: Thursday, July 5, 2018 Time: 11:00 Location: FMI-Building, Room 01.11.018

Learning to Hash for Large-Scale Medical Image Retrieval

Abstract:

The task of similarity search refers to fetching an item that is closest to the query item from a reference database under the notion of some distance measure. In critical applications of large-scale search and pattern matching, exhaustive comparison is often not possible due to prohibitive computational complexity and memory overheads. Towards mitigating this, hashing has been adopted as a popular approach for performing computationally efficient approximate nearest neighbor search. It is an approach of encoding data items into sequence of bits, called hash codes, such that the nearest neighbor search in the coding space is efficient and accurate.

With explosive growth in big data in medical imaging, there is a particular need for designing efficient indexing and search methods that can be leveraged for pattern exploration and knowledge discovery.

Towards learning optimized hashing functions for efficient indexing, we propose hashing forests (HF), which are an ensemble of code-efficient trees for recursive partitioning of the feature space. We propose two variants of HF, namely unsupervised HF (uHF) and metric HF (mHF) that are targeted at distance approximation and semantic-preserving hashing scenarios respectively. In uHF, trees are trained to parse the feature space into balanced well-clustered subspaces, whereas in mHF, the trees use localized metric learning to parse the space into semantically similar local neighborhoods. The versatility and efficacy of both these variants is demonstrated for the challenging task of pattern exploration within highly heterogeneous neuron databases. Extending to medical images, we leverage deep learning for end-to-end learning of convolutional neural networks for simultaneous representation learning and hashing. For enforcing code consistency within such networks, we design optimization objectives inspired by neighborhood component analysis (NCA) for minimizing the similarity weighted Hamming distance, with additional losses and regularizations such as bit-balance, bit-uncorrelation and quantization loss to improve hash code quality. We demonstrate the ability of such networks to perform semantics-preserving similarity search on a large scale chest X-ray image database with co-occurring disease patterns. We also investigate extension to multiple instance (MI) retrieval through the introduction of the MI pooling layer and robustness within NCA for improved learning. This was demonstrated on large-scale databases of breast mammography and histology for carcinoma assessment.

In conclusion, this thesis explores aspects of code-consistent training of hashing forests and deep learning models for end-to-end learning of hash codes and demonstrates that such hashing models can be leveraged to perform efficient and accurate large-scale content-based medical image retrieval.