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

• Published in: Scientific reports 2024-12, Vol.14 (1), p.30052-35
• Main Authors: Raju, Akella Subrahmanya Narasimha, Venkatesh, K., Padmaja, B., Kumar, CH. N. Santhosh, Patnala, Pattabhi Rama Mohan, Lasisi, Ayodele, Islam, Saiful, Razak, Abdul, Khan, Wahaj Ahmad
• Format: Article
• Language: English
• Subjects: 639/705/1042; 639/705/117; Accuracy; CKHK-22 dataset; Classification; Colorectal carcinoma; Colorectal Carcinoma (CRC); Colorectal Neoplasms - diagnosis; Databases, Factual; Datasets; Deep Learning; Diagnosis, Computer-Assisted - methods; Early Detection of Cancer - methods; Ensemble models; Humanities and Social Sciences; Humans; Image processing; Image Processing, Computer-Assisted - methods; Integrated CNNs; multidisciplinary; Neural networks; Neural Networks, Computer; Science; Science (multidisciplinary); Vision; Vision Transformers; XGBoost
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-024-81456-1

Early detection of colorectal carcinoma (CRC), one of the most prevalent forms of cancer worldwide, significantly enhances the prognosis of patients. This research presents a new method for improving CRC detection using a deep learning ensemble with the Computer Aided Diagnosis (CADx). The method involves combining pre-trained convolutional neural network (CNN) models, such as ADaRDEV2I-22, DaRD-22, and ADaDR-22, using Vision Transformers (ViT) and XGBoost. The study addresses the challenges associated with imbalanced datasets and the necessity of sophisticated feature extraction in medical image analysis. Initially, the CKHK-22 dataset comprised 24 classes. However, we refined it to 14 classes, which led to an improvement in data balance and quality. This improvement enabled more precise feature extraction and improved classification results. We created two ensemble models: the first model used Vision Transformers to capture long-range spatial relationships in the images, while the second model combined CNNs with XGBoost to facilitate structured data classification. We implemented DCGAN-based augmentation to enhance the dataset’s diversity. The tests showed big improvements in performance, with the ADaDR-22 + Vision Transformer group getting the best results, with a testing accuracy of 93.4% and an AUC of 98.8%. In contrast, the ADaDR-22 + XGBoost model had an AUC of 97.8% and an accuracy of 92.2%. These findings highlight the efficacy of the proposed ensemble models in detecting CRC and highlight the importance of using well-balanced, high-quality datasets. The proposed method significantly enhances the clinical diagnostic accuracy and the capabilities of medical image analysis or early CRC detection.