Deep Learning Algorithm to Determine the Presence of Rectal Cancer from Transrectal Ultrasound Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Diagnostic Technology

Background:

• Transrectal ultrasound (TRUS) is vital for rectal cancer detection, but accuracy depends on examiner experience.
• Deep learning, specifically convolutional neural networks (CNNs), offers potential to enhance medical image analysis.
• This study aimed to develop and evaluate a CNN for improved rectal cancer identification from TRUS images.

Purpose of the Study:

• To develop and assess a convolutional neural network (CNN) model for detecting rectal cancer using transrectal ultrasound (TRUS) images.
• To improve the diagnostic accuracy of rectal cancer identification, overcoming limitations related to examiner experience.
• To provide a computational tool that assists clinicians in diagnosing rectal cancer.

Main Methods:

• Retrospective collection of 681 TRUS images (August 2008 - September 2022).
• Image classification into 'rectal cancer' and 'normal rectum' categories.
• Training a CNN model utilizing the EfficientNetV2-S architecture for image differentiation.

Main Results:

• The CNN model achieved high performance metrics: 96.7% training accuracy and 90.5% validation accuracy.
• Area Under the Curve (AUC) values were 0.996 (training) and 0.945 (validation).
• Specific performance for rectal cancer: precision 0.935, recall 0.944, F1-score 0.940; for normal rectum: precision 0.793, recall 0.767, F1-score 0.780.

Conclusions:

• The developed CNN model demonstrates strong capability in differentiating rectal cancer from normal rectum in TRUS images.
• The model serves as a valuable decision-support tool for clinicians, potentially enhancing diagnostic workflows.
• Further research is recommended to broaden the model's generalizability and incorporate stage classification.