Automatic detection and recognition of lesions in medical images through artificial intelligence algorithms provides an economic, efficient, and effective computer-aided approach for clinical diagnosis, treatment, and prognosis monitoring. 

However, automatic lesion (organ or tissue) recognition in medical images still remains challenging: the sizes and shapes of lesion regions vary among individuals; in some cases, obvious individual differences increase the recognition difficulty; high-precision segmentation is prone to the low contrast between the lesions and background.  

To solve the challenges of medical Image segmentation, a research team led by Prof. FAN Jianping and LI Ye from the Shenzhen Institute of Advanced Technology (SIAT) of Chinese Academy of Sciences has proposed a boundary-aware context neural network.  

Their research was published in Medical Image Analysis. 

Benefiting from the designed pyramid edge extraction module, multi-task learning module, and cross-feature fusion module, multi-level and fine-grained image features were adaptively extracted. This improved the perception of the neural network on complex structures such as lesion shape, distribution, and edge information, and reduced the interference of surrounding normal tissues, organs, and noise.

The researchers have quantitatively and qualitatively validated the proposed method in various lesion segmentation tasks of multi-modal medical images, such as dermoscopy images, endoscope images, and X-ray images.

Compared with other deep learning methods, this approach exhibited better performance. Concretely, the recognition accuracy for the melanoma segmentation based on dermoscopy image was 81.0%; that for colon polyp recognition based on endoscope image was 88.5%; that for lung organ segmentation based on X-ray image was 92.8%.

"The target areas are accurately located with the aid of effective image context representations. This indicates that our model is able to simultaneously process fine structures and rectify errors, which can assist doctors to speed up the diagnosis process and improve diagnosis precision," said Dr. WANG Ruxin, the first author of this study.