Localization of Colorectal Cancer Lesions in Contrast-Computed Tomography Images via a Deep Learning Approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Abdominal computed tomography (CT) is crucial for gastrointestinal disease evaluation, including colorectal cancer detection.
• Analyzing the colon in CT scans can be challenging due to its structure, potentially leading to missed incidental colon cancers.
• Current CT sensitivity for colorectal cancer detection is tumor-size dependent, highlighting the need for improved diagnostic tools.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model for the automatic localization of colorectal cancer in CT images.
• To address the challenge of detecting incidental colon cancers in CT scans of unprepared bowels.
• To improve the diagnostic accuracy and efficiency of colorectal cancer screening using CT imaging.

Main Methods:

• Utilized a dataset of 1558 tumor slices from 190 colorectal cancer patients, annotated by radiologists and surgeons.
• Employed and compared deep learning models including RetinaNet, YOLOv3, and YOLOv8 for tumor localization.
• Validated tumor sites through colonoscopy, physical examination, image studies, and operation records.

Main Results:

• The YOLOv8 model, with hyperparameter tuning, demonstrated strong performance in patient-wise testing.
• Achieved a sensitivity of 0.83 (±0.29), specificity of 0.97 (±0.01), and accuracy of 0.96 (±0.01).
• Slice-wise testing yielded an F1 score of 0.97 (±0.002) and a mean Average Precision (mAP) of 0.984.

Conclusions:

• AI, specifically the YOLOv8 model, shows significant potential for accurately localizing colorectal cancer in CT images.
• This AI-driven approach can aid in the detection of incidental colon cancers, improving patient outcomes.
• The study underscores the value of AI in enhancing the diagnostic capabilities of CT for colorectal cancer screening.