Cervical cancer screening using a deep learning model of precancerous colposcopy images in Resource constrained Areas,2024

Abstract

Introduction: Cervical cancer remains a significant global health challenge, particularly in resource-constrained settings where access to effective screening programs is limited. Traditional screening methods face limitations in these contexts, leading to delayed diagnoses and poor patient outcomes. Deep learning techniques offer a promising solution by leveraging advanced algorithms to analyze colposcopy images for early detection of precancerous lesions. Objective: This study aims to develop and validate cervical cancer screening using deep learning with precancerous image, 2024. Methods: It involves a cross-sectional study the data from the International Agency for Research on Cancer (IARC) containing colposcopy examination were 902 Images accessed. A deep learning approach was employed for cervical cancer screening using colposcopy images with labels for LSIL, HSIL, and Normal cervix. After manually curating the data set it was split into training and validation sets, followed by data augmentation with random transformations. Transfer learning was utilized, incorporating models such as InceptionResNetV2, InceptionV3, VGG19, and ResNet152V2 for feature extraction, with fine-tuning applied for three-class classification. A two-layer neural network with dropout and ReLU activation was added, using categorical cross-entropy as the loss function and the Adam optimizer. Hyperparameter tuning (batch size, learning rate, and regularization) was performed, and the model's performance was evaluated using metrics such as accuracy, precision, recall and AUC. Result: In this study, a total of 527 cervical images were used after removing images with excessive background noise and green light filters, consisting of 194 HSIL, 48 LSIL, and 285 normal cervical images. The images were cropped to 60% along the X and Y axes and split using an 80-20 rule into training and test sets. The training set was augmented and balanced, resulting in 966 images per class (2,898 total). The test set was also augmented, reaching 584 images. InceptionResNetV2, with frozen layers and a custom top layer, achieved the highest performance across all metrics with a validation accuracy of 82.79%, precision of 82.99%, recall of 82.27%, and an AUC of 94.81. The model's loss decreased consistently over the epochs without overfitting. A confusion matrix demonstrated strong classification performance. Grad-CAM visualizations further highlighted the model's focus on regions of interest in the images. Conclusion: This study developed and validated a deep learning model for cervical cancer screening using precancerous colposcopy images, achieving an accuracy of 82.3% with the InceptionResNetV2 transfer learning model. The findings highlight the feasibility of using automated screening techniques in low-resource settings where specialized human resources for colposcopy reading are limited, demonstrating the potential of deep learning and AI to provide objective, cost-effective screening and diagnostic solutions in healthcare.