Base on neuroendocrine tumour classification in CT imaging using pretrained vision transformers and data augmentation

#4586

Introduction: Neuroendocrine tumours (NETs) are a group of cancers that can be difficult to distinguish from other malignancies in CT imaging. Their diverse appearance often leads to diagnostic challenges, requiring advanced techniques to improve classification accuracy. Pretrained Vision Transformers (ViTs) have shown promise in medical image analysis, but additional methods are needed to enhance model robustness, particularly when dealing with complex and heterogeneous tumour characteristics.

Aim(s): This study investigates the use of pretrained Vision Transformers for classifying neuroendocrine tumours in CT imaging, with a focus on improving performance through data augmentation.

Materials and methods: We used a dataset of CT scans containing annotated NETs and other cancer types, such as pancreatic cancers. A pretrained Vision Transformer model was fine-tuned on this dataset to identify distinguishing features of NETs. To enhance the model’s generalisation and robustness, data augmentation techniques were employed, which helped simulate a variety of tumour appearances and reduce overfitting. Model performance was evaluated using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC).

Conference:

Presenting Author:

Authors: Li H, Tang Z, Ding Z, Chen Y, Li H,

Keywords: Vision transformers, Cancer Detection, Transfer Learning, Medical Image Analysis,