The following is a summary of the “Deep Learning-Based Model for Identifying Tumors in Endoscopic Images From Patients With Locally Advanced Rectal Cancer Treated With Total Neoadjuvant Therapy,” published in the March 2023 issue of Hematology by Thompson, et al.
The purpose of this research was to create and evaluate a novel method for automatically classifying endoscopic pictures for the presence or absence of a tumor using a deep convolutional neural network. Endoscopic images were collected before, during, and after entire neoadjuvant therapy, and were categorized into groups according to tumor presence in this prospective pilot investigation. An endoscopic picture set was used to train a modified convolutional neural network for probabilistic classification of tumor vs no tumor. After the network was trained, it was put through its paces with a testing set of endoscopic images. The research was performed at a hospital specializing in cancer care.
They looked at images from 109 patients who were given entire neoadjuvant therapy after being diagnosed with locally advanced rectal cancer between December 2012 and July 2017. The primary result was the percentage of test photos where a tumor was correctly identified as present or absent (area under the receiver operating characteristic). There were a total of 1,392 images used, with 1,099 used for training (468 benign and 631 malignant) and 293 used for testing (151 benign and 142 malignant).
The combined training/testing area under the receiver operating characteristic was 0.83. The investigation was conducted at a single location and only a small number of photos were collected from each group. Moderate accuracy is achieved when using a convolutional neural network to identify tumors from healthy tissue. The convolutional neural network should be tested on a huge dataset of images in future studies.
