Abstract
BACKGROUND: Transrectal ultrasound (TRUS) is a crucial diagnostic tool for accurately detecting rectal cancer; however, its accuracy varies with the examiner's experience. Deep learning, particularly convolutional neural networks (CNNs), exhibited promise in improving diagnostic accuracy in medical imaging. This study developed and assessed a CNN model for identifying rectal cancer using TRUS images. METHODS: We retrospectively gathered 681 TRUS images that were obtained between August 2008 and September 2022. The images were classified as rectal cancer and normal rectum. Then, a CNN model was trained using the EfficientNetV2-S architecture to differentiate between rectal cancer and normal rectum images. RESULTS: Of the 681 TRUS images, 533 and 148 were obtained from rectal cancer and normal rectum cases, respectively. The CNN model achieved training and validation accuracies of 96.7% and 90.5% and areas under the curve of 0.996 and 0.945, respectively. The precision, recall, and F1 scores were 0.935, 0.944, and 0.940 for rectal cancer and 0.793, 0.767, and 0.780 for the normal rectum, respectively. CONCLUSIONS: Our CNN model exhibited good performance in distinguishing rectal cancer from the normal rectum in TRUS images. The model is a valuable decision-support tool to help clinicians. Future studies are warranted to improve the model's generalizability and enable stage classification integration.