Uterine cancer classification from CT images using convolutional feature extraction and transformer-based self-attention

TL;DR

This paper introduces a new AI model that combines convolutional and transformer techniques to improve the accuracy of diagnosing uterine cancer from CT scans.

Contribution

The novel hybrid framework integrates convolutional feature extraction with transformer-based global attention for uterine cancer classification.

Findings

The proposed model achieves 87.44% accuracy and 99.41% AUC in classifying uterine cancer from CT images.

It outperforms existing deep learning models like VGG16, ResNet50, and MobileNetV2 in classification metrics.

The integration of DenseNet121 and transformer mechanisms enhances both local and global feature representation.

Abstract

Accurate and early diagnosis of uterine cancer from computed tomography images remains a challenging task due to the complexity of anatomical structures and the subtle visual differences between normal, benign, and malignant uterine tissues. Traditional diagnostic approaches and conventional deep learning models often fail to effectively capture both local and global image characteristics.

This study aims to develop and validate a novel hybrid deep learning framework that integrates convolutional feature extraction with transformer-based global attention mechanisms to improve the accuracy and robustness of uterine cancer classification from computed tomography images.

In the proposed framework, DenseNet121 is employed as a convolutional neural network feature extractor, while a transformer encoder is utilized to model long-range contextual dependencies through multi-head self-attention. DenseNet121 captures discriminative local features from computed tomography images, which are subsequently processed by the transformer to enhance global feature representation. The performance of the proposed model is evaluated using the KAUH uterine cancer computed tomography dataset, which includes three classes: normal, benign, and malignant. The proposed approach is compared with several state-of-the-art deep learning models, including VGG16, VGG19, MobileNetV2, ResNet50, and DenseNet121.

Experimental results demonstrate that the proposed hybrid model outperforms the comparative models. It achieves an accuracy of 87.44%, sensitivity of 87.13%, specificity of 95.20%, an F1 score of 87.17%, and an area under the receiver operating characteristic curve of 99.41%.

The results confirm the effectiveness of integrating convolutional neural networks with transformer-based self-attention mechanisms for significantly improving uterine cancer classification from computed tomography images. The proposed model shows strong potential as a computer-aided decision-support tool for radiologists to assist in the detection of uterine cancer and may be extended to various real-world clinical applications.

Figures