Transfer Recurrent Feature Learning for Endomicroscopy Image Recognition

by

Transfer Recurrent Feature Learning for Endomicroscopy Image Recognition

The Hamlyn Centre researchers proposed a transfer recurrent feature learning framework for classification tasks on pCLE videos to aid cancer diagnosis

Probe-based confocal laser endomicroscopy (pCLE) is an emerging tool for epithelial cancer diagnosis, which enables in-vivo microscopic imaging during endoscopic procedures and facilitates the development of automatic recognition algorithms to identify the status of tissues.

Using flexible fiber-bundle and miniaturized optics, it provides clinicians with real-time access to histological information during surgical procedures and has demonstrated promising sensitivities and specificities in various pre-clinical and clinical studies, including in the gastro-intestinal tract, breast and lungs.

Although pCLE enables the acquisition of in-vivo microscopic images that resemble closely to histology images, reliable diagnosis is still challenging for many clinicians who have little histopathology expertise and training.

Furthermore, the high variability in the appearances of pCLE images and the presence of atypical conditions makes it difficult to provide accurate diagnosis by manual identification.

A Transfer Recurrent Feature Learning (TRFL) Framework

The researchers at the Hamlyn Centre proposed a Transfer Recurrent Feature Learning (TRFL) framework for classification tasks on pCLE videos, aiming to aid the probe-based confocal laser endomicroscopy (pCLE) recognition.

A transfer recurrent feature learning framework
The main workflow of the proposed framework

At the first stage, the discriminative feature of single pCLE frame was learned via generative adversarial networks based on both pCLE and histology modalities.

Synthetic data of pCLE and histology patches. The left side of figure is the conversion ‘Histology→ pCLE→ Histology’ and the right side is ‘pCLE→ Histology→ pCLE’. (a) Real Histology Patch. (b) Synthesized pCLE Frame. (c) Reconstructed Histology Patch. (d) Real pCLE Frame. (e) Synthesized Histology Patch. (f) Reconstructed pCLE Frame.
Synthetic data of pCLE and histology patches. The left side of figure is the conversion ‘Histology→ pCLE→ Histology’ and the right side is ‘pCLE→ Histology→ pCLE’. (a) Real Histology Patch. (b) Synthesized pCLE Frame. (c) Reconstructed Histology Patch. (d) Real pCLE Frame. (e) Synthesized Histology Patch. (f) Reconstructed pCLE Frame.

At the second stage, our researchers used recurrent neural networks to handle the vary length and irregular shape of field of view in pCLE mosaics, taking the frame-based features as input.

The experiments based on real pCLE data sets demonstrated that our approach outperforms, with statistical significance, state-of-the-art approaches. A binary classification accuracy of 84.1% has been achieved.

synthetic pCLE mosaic
(a) is the subregion of histology slides; (b) is the zoom out of region with red rectangle; (c) is the synthetic pCLE mosaic generated from histology; (d) is the corresponded region of (b).

FIND OUT MORE>>


This research was supported by EPSRC Programme Grant “Micro-robotics for Surgery (EP/P012779/1)” and was published in IEEE Transactions on Medical Imaging (Yun Gu; Khushi Vyas; Jie Yang and Guang-Zhong Yang, "Transfer Recurrent Feature Learning for Endomicroscopy Image Recognition").

Supporters

Reporter

Erh-Ya (Asa) Tsui

Erh-Ya (Asa) Tsui
Enterprise

Click to expand or contract

Contact details

Tel: +44 (0)20 7594 8783
Email: e.tsui@imperial.ac.uk

Show all stories by this author

Tags:

Lung-disease, Cancer, Imaging, Research, Surgery
See more tags