Applying Deep Transfer Learning to Assess the Impact of Imaging Modalities on Colon Cancer Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate colon cancer detection is crucial, with medical imaging playing a key role.
• Deep learning (DL) methods' performance depends on the quality of medical images.
• Effective imaging modalities for DL-based colon cancer detection require comprehensive evaluation.

Purpose of the Study:

• To comprehensively report the performance of various imaging modalities with DL models for colon cancer detection.
• To identify the optimal imaging modality and DL model combination for colon cancer detection using transfer learning (TL).
• To compare the efficacy of different DL architectures and ensemble models.

Main Methods:

• Utilized three imaging modalities: computed tomography, colonoscopy, and histology.
• Employed five DL architectures: VGG16, VGG19, ResNet152V2, MobileNetV2, and DenseNet201.
• Assessed models using 5400 images (normal and cancerous) on a high-performance GPU, including ensemble models.

Main Results:

• Colonoscopy with the DenseNet201 model achieved the highest average performance of 99.1% (AUC, precision, F1).
• This combination outperformed all other individual and ensemble DL models evaluated.
• The study provides a detailed comparison of imaging modalities and DL models in a TL setting.

Conclusions:

• Colonoscopy is the most effective imaging modality for DL-based colon cancer detection.
• DenseNet201 is the optimal DL model when used with colonoscopy images for this task.
• The findings guide research organizations in selecting effective imaging and DL strategies for colon cancer detection.