Automatic anatomical classification of colonoscopic images using deep convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colonoscopy is crucial for detecting colorectal diseases like cancer and polyps.
• Computer-aided diagnosis (CAD) systems using deep convolutional neural networks (CNNs) can aid practitioners by recognizing anatomical locations during colonoscopies.
• This study aimed to develop a CNN-based CAD system for distinguishing colorectal anatomical regions.

Purpose of the Study:

• To construct a CNN model capable of classifying colonoscopic images into specific anatomical locations.
• To evaluate the performance of the CNN in recognizing seven distinct anatomical regions of the colon.

Main Methods:

• A CNN was trained on 9,995 colonoscopy images and validated on 5,121 independent images.
• Images were categorized into seven anatomical locations: terminal ileum, cecum, ascending to transverse colon, descending to sigmoid colon, rectum, anus, and indistinguishable areas.
• The study evaluated the concordance between CNN diagnoses and endoscopist diagnoses, focusing on sensitivity and specificity.

Main Results:

• The CNN achieved high areas under the curve (AUC) for anatomical recognition: 0.979 (terminal ileum), 0.940 (cecum), 0.875 (ascending-transverse colon), 0.846 (descending-sigmoid colon), 0.835 (rectum), and 0.992 (anus).
• The CNN system demonstrated a 66.6% overall correct recognition rate on the test dataset.
• The CNN showed clinically relevant performance in identifying anatomical landmarks.

Conclusions:

• A novel CNN system was developed with clinically relevant performance for anatomical recognition in colonoscopy images.
• This represents a significant first step towards a CAD system that can support colonoscopy procedures.
• The developed system has the potential to enhance the quality assurance of colonoscopies.