Uterine Cancer Image Analysis with Convolutional Neural Networks

Endometrial uterine cancer is the fourth most common cancer in women in the United States. Unlike most cancers, its rates have rising incidences and mortalities. Although Black women are less likely to be diagnosed with the disease, they are more likely to die from it. This discrepancy can be partially credited to the fact that Black women are often diagnosed at a later, less treatable stage. Studies attribute this diagnostic discrepancy to a variety of factors, including a lack of accurate and precise diagnosis during early stages of the cancer. A possible solution is using artificial intelligence and machine learning (AI/ML) models. Machine learning techniques are shown to be excellent at diagnosing cancer through the analysis of data and images such as X-rays, magnetic resonance imaging (MRIs), and ultrasounds. However, a major limitation of machine learning techniques is that if the models are trained on bad or biased data, they will reproduce that bias in their evaluations. Similarly, if diagnostic devices, such as sensors, are designed without an understanding of cell behavior differences based on ethnicity, the device will also have biases in its outcomes. Convolutional neural networks (CNNs) are an advanced form of ML that show promise in medicine, accurately performing diagnoses and medical image analysis. A key advantage of CNNs is less required manual image preprocessing compared to other ML techniques. Although, CNNs require large datasets to be trained from scratch, repurposing pre-trained CNNs and artificially expanding the dataset significantly cut down the required data collection. To address racial disparities in patient survival rates, this study developed convolutional neural network models that predict type of endometrial cancer in medical images, using pre-trained machine learning algorithms and ultrasound images. Analyzing the change in the accuracy of models as a function of race will determine if there is a bias.CNNs - ResNet50, DenseNet-201, VGG16, AlexNet, and Inception ResNet-V2 - were able to discern the race of the patient and the type of cancer with accuracies of 98.31%, 97.06%, 97.92%, 93.18, and 96.67% respectively, demonstrating the power of CNNs in medical diagnosis. Training the model on a racially biased dataset led to a notable decrease in test accuracy on the underrepresented races. ResNet50 could discern the type of cancer 86.42% of the time when there were an equal number of black and white patients in the training set, but the accuracy dropped to 61.22% when trained on only white patients. The other models displayed a similar drop-off in performance. This signifies a need to ensure a racially diverse training set to avoid introducing biases into the diagnostic process. Using the source data, we were able to 1) identify machine learning models that can recognize uterine cancer through the analysis of existing scans at high accuracies; and 2) identify racial disparities in diagnosing endometrial uterine cancer via ultrasound imaging and analysis. The methodology used by this study is applicable to other diseases diagnosed with medical imaging and can assist with identifying and solving racial disparities.