Deep-Learning Approach Precisely Identifies Potentially Cancerous Growth in Colonoscopy Images

Colonoscopy is the established method for detecting colorectal growths or 'polyps' in the inner lining of the colon, which can lead to rectal cancer if left untreated. Through the analysis of colonoscopy images, medical professionals can identify polyps early and prevent further complications. This process involves "polyp segmentation," distinguishing polyp segments from normal layers of colon tissue. While traditionally performed by humans, computer algorithms utilizing deep learning have made significant progress in polyp segmentation.

However, two main challenges persist. The first challenge involves image "noise" caused by rotational movements of the colonoscope lens during image capture, leading to motion blur and reflections. This blurs the boundaries of polyps, making segmentation difficult. The second challenge is the natural camouflage of polyps, as their color and texture often resemble surrounding tissues, resulting in low contrast. This similarity hampers accurate polyp identification and adds complexity to segmentation.

To address these challenges, researchers from Tsinghua University (Beijing, China) have developed two modules to enhance the use of artificial neural networks for polyp segmentation. The "Similarity Aggregation Module" (SAM) addresses rotational noise issues by extracting information from individual pixels and using global semantic cues from the entire image. Graph convolutional layers and non-local layers are employed to consider the mathematical relationships between all parts of the image. The SAM achieved a 2.6% performance increase compared to other state-of-the-art polyp segmentation models.

To tackle camouflage difficulties, the "Camouflage Identification Module" (CIM) captures subtle polyp clues concealed within low-level image features. The CIM filters out irrelevant information, including noise and artifacts, which could interfere with accurate segmentation. With the integration of the CIM, the researchers achieved an additional 1.8% improvement in performance. The researchers now aim to optimize the method by implementing techniques like model compression, reducing computational complexity for practical use in real-world medical settings.

Related Links:
Tsinghua University