תלמידת המחלקה להנדסה ביו רפואית לתואר שני תרצה בנושא:

Lung Texture Classification Using Bag of Visual Words

Interstitial lung diseases (ILD) are a group of more than 150 disorders of the lung parenchyma. High-Resolution Computed Tomography (HRCT) is the most important modality in ILD diagnosis. Classification of various lung tissue patterns caused by ILD is considered a very challenging task. The main difficulties are that different lung textures look very similar and there are various texture patterns representing a specific tissue type.

This study focuses on classification of five most common categories of lung tissues of ILD on HRCT images:  normal, emphysema, ground glass, fibrosis and micronodules. Our objective is to classify an annotated region of interest (AROI) using a bag of visual words (BoVW) framework. We evaluate different feature descriptors extracted from each image. A collection of representative feature vectors are defined as visual words and form a dictionary. Dictionary construction can be defined for the entire image set, as well as individually per lung texture category. We compare both approaches. The classification is performed using an SVM classifier with histogram intersection kernel. Additionally, we perform the classification on small (31x31) ROIs within the AROIs - and compute a max vote from the individual ROIs. The influence of this voting algorithm on the classification performance is investigated. In order to estimate the generalization performance of the classification of AROI, a leave-one-patient-out cross validation (LOPO CV) is used. Dataset of 1018 AROIs from 95 patients is used. We experiment with various parameters in order to optimize the system performance. Our classification results compete with the state-of-the-art in this domain.

~~Speaker: Ayelet Aharon,
M.Sc. student under the supervision of Prof. Simon Litsyn

Wednesday, March 11, 2015 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering

Adaptive Successive Cancellation List Decoding of Polar Codes

Abstract
Polar codes, invented by Arikan, are the first to achieve the Shannon capacity over discrete memoryless channels (DMC) with feasible implementation complexity of O(n*log(n)), where n is the code's length. Successive Cancelation List (SCL) decoding combined with Cyclic Redundancy Check (CRC) improves polar codes' performances to the extent of being better than those of modern coding techniques. The SCL decoder explores simultaneously L decoding paths which results in running time complexity of O(L*n*log(n)) and space complexity of O(L*n) where L is the list size.

In this work an adaptive version of the SCL decoder, namely ASCL decoder, is proposed. While decoding, the ASCL decoder decides, with the help of predefined thresholds, how many paths to allocate at each decoding step according to the input's noise level and the vulnerability of the currently decoded bit to mistakes. At each decoding step, up to L_max paths can be allocated (if a threshold has isolated more paths the algorithm uses sorting to detect the L_max most probable paths) leading to a space complexity of O(L_max*n). The running time complexity is O(L_average*n*log(n)), where L_average≤L_max is the average list size the decoder uses, which can be monitored by changing the thresholds suitably.

For a given running time complexity, the proposed algorithm's BLER performance is significantly improved in comparison to this of the SCL decoder. Furthermore, for a high enough SNR, the BLER performance of the ASCL decoder with any 2≤L_average≤L_max  is similar to this of the SCL decoder with list size L=L_max. This makes the ASCL decoder a valuable replacement to the SCL decoder.

Unlike the SCL which requires to perform sorting at (almost) each decoding step, the ASCL decoder avoids most of these procedures due to its use of thresholds. By utilizing this property even further, we suggest modified versions of the ASCL decoder which completely avoid sorting at the expense of a relatively small decrease of the BLER performance.

Finally, since the ASCL decoder (as well as any other list decoding procedure), relies on the use of a CRC, we establish a set of guidelines to assist in finding the optimal CRC length for a given SNR value. A wise choice of the CRC can prevent an unnecessary and sometimes significant loss in performance, especially when the decoder's performance has the potential of improving drastically as the SNR grows.