GPTNeXt: Biomedical Image Classification Investigations.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Imaging

Background:

• Prominent computer vision solutions utilize transformers and convolutional neural networks (CNNs).
• Integrating CNNs and transformers is common for image classification model development.

Purpose of the Study:

• Introduce an innovative CNN model, GPTNeXt, inspired by the Generative Pretrained Transformer (GPT) architecture.
• Assess the image classification capabilities of the proposed GPTNeXt model on diverse biomedical datasets.

Main Methods:

• Developed the GPTNeXt model and a deep feature engineering approach.
• Extracted features using a novel method from global average pooling and nine fixed-size image patches.
• Applied iterative neighborhood component analysis (INCA) for feature selection and used three shallow classifiers for classification.

Main Results:

• The GPTNeXt-based feature engineering model achieved over 98% classification accuracy on Alzheimer's disease, blood, and lung cancer image datasets.
• Demonstrated versatile classification performance across multiple biomedical imaging tasks.

Conclusions:

• The proposed GPTNeXt approach is highly effective for biomedical image classification.
• A lightweight CNN variant also demonstrated outstanding classification performance, highlighting model efficiency.