Exploring vision transformers and XGBoost as deep learning ensembles for transforming carcinoma recognition

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal carcinoma (CRC) is a leading cause of cancer worldwide, making early detection crucial for improved patient outcomes.
• Current diagnostic methods face challenges with imbalanced datasets and complex feature extraction in medical image analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning ensemble method for enhanced early detection of colorectal carcinoma (CRC).
• To address challenges of imbalanced datasets and improve feature extraction in medical image analysis for CRC detection.

Main Methods:

• Utilized pre-trained convolutional neural network (CNN) models (ADaRDEV2I-22, DaRD-22, ADaDR-22) combined with Vision Transformers (ViT) and XGBoost for a deep learning ensemble.
• Refined the CKHK-22 dataset from 24 to 14 classes to improve data balance and quality, enabling more precise feature extraction.
• Implemented DCGAN-based augmentation to enhance dataset diversity and created two ensemble models: one with ViT for spatial relationships, another with CNNs and XGBoost for structured data.

Main Results:

• The ADaDR-22 + Vision Transformer ensemble achieved the highest performance with 93.4% testing accuracy and 98.8% AUC.
• The ADaDR-22 + XGBoost model demonstrated strong results with 92.2% accuracy and 97.8% AUC.
• The refined dataset and ensemble approach led to significant improvements in classification accuracy and feature extraction.

Conclusions:

• The proposed deep learning ensemble models are highly effective for early colorectal cancer detection.
• The study underscores the importance of well-balanced, high-quality datasets in medical image analysis for improving diagnostic accuracy.
• This method significantly enhances clinical diagnostic accuracy and capabilities in medical image analysis for early CRC detection.