Browse Prior Art Database

Lesion detection in medical images by cascade classification of bag of boundary features Disclosure Number: IPCOM000237626D
Publication Date: 2014-Jun-27
Document File: 8 page(s) / 66K

Publishing Venue

The Prior Art Database


This paper presents a novel method for automatic lesion detection in breast ultrasound images; the method performs multi-stage learning of lesion-specific boundaries represented by a bag of robust features. The proposed method can be seen as an edge pruning procedure that leaves only object-specific edges and fillters out the rest. It can be combined with segmentation algorithms that rely on edge information. We show an example of such combination with one of the state-of-art segmentation algorithms; our method yields improved segmentation results. The proposed method is tested on a set of 400 breast ultrasound images, with the goal to automatically detect lesion boundaries. However, we believe that our method can be used by radiologists as an assistance tool during examination routine, in which case it may help to better localize lesions and document the findings. The performance of our method is compared to a state-of-art object boundary classification algorithm; we show that our method outperforms it in di erent tests.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 12% of the total text.

Page 01 of 8

Lesion detection in medical images by cascade classification of bag of boundary features

Lesion detection in medical images is an active area of research. Lesion detection methods rely on object boundary detection in one way or another. The most common method for this step of edge detection is based on the Canny edge detector, which usually provides consistent results. The drawback is that this method does not take into account specific properties of lesions that are to be detected. On the other hand, one of the the current state-of-art in the edge detection is the probability of boundaries contour detection [1] which is trained to detect edges in natural images, and then to combine the edges into segments. This approach uses local features calculated from the patch around pixel on the boundary. This approach does not take into account shape properties of an object to be detected and only find most probable edge pixels separate from each other.

Furthermore, object boundaries in natural images are usually well defined, and, therefore, objects can be clearly distinguished from their background. In contrast, in medical images, objects such as lesions or even organs do not always have fully closed contours with clear boundaries. The reasons for that include limitations of acquisition methods and devices (for example, having low resolution and/or low dynamic range), occlusions of other tissues and organs, and others. In these cases, methods like [1] may find boundaries that have no physical meaning. In particular, such methods merge neighboring areas if they are 'similar' enough, or if there are no sufficiently strong boundaries detected between the areas, and miss the actual lesion boundaries.

We present a novel method for automatic lesion detection in medical images. The proposed method performs edge pruning by means of supervised learning.

The method is based on a new multi-stage algorithm for learning lesion-specific boundaries represented by a bag of robust features. New ideas include: 1) local boundary classification decision is dependent on global features, 2) separate boundary segments are learnt to be merged in a multigrid supervised manner, 3) we start from smallest scale and propagate the classification decision to larger scales.

The proposed ideas allow the detection to adapt automatically to a specific type of lesion to be detected, to detect low contrast boundaries, and to preserve the natural shape of the lesion.

In the following description of our new method for lesion detection, we provide implementation details.

We first identify separate edge sections where each pixel (or, alternatively, each m-th pixel) is a center of the local neighborhood that has
a disk shape. We then calculate a local area descriptor for each one of the pixel areas. Such descriptor can include different features including single or pairwise local descriptors, texture, shape, intensity based etc.

The next step is the the 1st training classificati...