Scattering features for lung cancer detection in fibered confocal fluorescence microscopy images

作者：Rakotomamonjy Alain^*; Petitjean Caroline; Salauen Mathieu; Thiberville Luc

来源：Artificial Intelligence in Medicine, 2014, 61(2): 105-118.

DOI：10.1016/j.artmed.2014.05.003

• Scattering transform
• Wavelet transform
• Local binary pattern
• Support vector machine
• Texture analysis
• Fibered confocal fluorescence microscopy imaging
• Bronchoscopy

Objective: To assess the feasibility of lung cancer diagnosis using fibered confocal fluorescence microscopy (FCFM) imaging technique and scattering features for pattern recognition. %26lt;br%26gt;Methods: FCFM imaging technique is a new medical imaging technique for which interest has yet to be established for diagnosis. This paper addresses the problem of lung cancer detection using FCFM images and, as a first contribution, assesses the feasibility of computer-aided diagnosis through these images. Towards this aim, we have built a pattern recognition scheme which involves a feature extraction stage and a classification stage. The second contribution relies on the features used for discrimination. Indeed, we have employed the so-called scattering transform for extracting discriminative features, which are robust to small deformations in the images. We have also compared and combined these features with classical yet powerful features like local binary patterns (LBP) and their variants denoted as local quinary patterns (LQP). %26lt;br%26gt;Results: We show that scattering features yielded to better recognition performances than classical features like LBP and their LQP variants for the FCFM image classification problems. Another finding is that LBP-based and scattering-based features provide complementary discriminative information and, in some situations, we empirically establish that performance can be improved when jointly using LBP, LQP and scattering features. %26lt;br%26gt;Conclusions: In this work we analyze the joint capability of FCFM images and scattering features for lung cancer diagnosis. The proposed method achieves a good recognition rate for such a diagnosis problem. It also performs well when used in conjunction with other features for other classical medical imaging classification problems.

• 出版日期2014-6