DEEP LEARNING IN ENDOMETRIAL CANCER: PREDICTING MALIGNANCY FROM HISTOPATHOLOGICAL IMAGES OF QATARI PATIENTS.

Abstract

Endometrial cancer is the second most common gynecological malignancy, and an accurate diagnosis is essential for determining appropriate treatment and improving patient outcomes. Traditional diagnostic methods rely on histopathological examination, which can be subjective and prone to variability. Deep learning techniques have emerged as powerful tools to automate and enhance diagnostic workflows in pathology. This study assesses the performance of a deep learning model in predicting malignancy in endometrial cancer using digitized hematoxylin and eosin (H&E) stained histopathological slides from Qatari patients. A retrospective analysis was conducted on 1,312 slides from 70 patients, which were scanned using a slide scanner to obtain high-quality digitized images. Through a collaborative effort with expert pathologists, we classified 880 images as benign or malignant, which were then divided into training (60%, 527 images), validation (20%, 176 images), and test (20%, 177 images) sets. Our study employed the deep learning model-MobileNetV2, a convolutional neural network (CNN)-to classify the tissue samples. The model's training accuracy improved from 75% at epoch 0 to 95% by epoch 50. On the test set, the model achieved an overall accuracy of 90.96%, with a precision of 90.22%, a recall of 92.22%, and an F1-score of 91.21%. The confusion matrix revealed that 83 out of 90 malignant cases and 78 out of 87 benign cases were correctly classified, with 9 false positives and 7 false negatives. These detailed quantitative results underscore the effectiveness of our deep learning approach in accurately distinguishing between benign and malignant endometrial tissues, highlighting its potential for integration into clinical practice to enhance diagnostic precision and support clinical decision-making. Moreover, this method can assist in prioritizing cases for review, ensuring that high-risk samples receive immediate attention, which is critical for timely patient management. Importantly, by using locally sourced data from Qatari patients, our findings lay the foundation for developing a region-specific deep learning model that captures the unique histopathological nuances present in this population. Such a model could further optimize diagnostic accuracy and ultimately contribute to improved patient outcomes in Qatar, providing a valuable tool for personalized and efficient cancer care in the future.