The use of artificial intelligence in MRI diagnostics of rectal cancer

Abstract

AIM: to work out decision-making support systems for MRI diagnostics of rectal cancer: site and segmentation of the primary tumor. PATIENTS AND METHODS: the study included 450 MRI studies of patients with rectal cancer and 450 MRI studies of patients without. All patients with tumors of rectum had histological verification. Data were collected in T2Wcoronal and axial projections (MRI Philips Achieva 1.5 T). Object marking was carried out only for T2W projections, where the area of interest was segmented — rectum, sigmoid colon and tumor. The ITK-Snap program was used to label MRI images. The validated studies and labeling were used to create a machine learning model that demonstrates the capability of the dataset to build medical decision support systems. SegResNet, Trans Unet, 3D Unet neural networks were used to create a basic artificial intelligence model. RESULTS: dice similarity coefficient (DSC) of various neural networks were: TransUnet — 0.33, SegResNet — 0.50, 3D Unet — 0.42. The diagnostic efficiency of the SegResNet neural network in detecting rectal tumors with the addition of negative examples and post-processing was accuracy 77.0%; sensitivity 98.1%; specificity 45.1%; positive predictive value 72.9%; negative predictive value of 94.1%. At this stage, AI has a high sensitivity and accuracy, which indicates a high diagnostic efficiency in terms of visualizing the primary tumor and determining localization in the rectum. However, the specificity of the method is still at an unsatisfactory level (45.1%), which indicates a high percentage of false positive results in healthy patients and does not allow the model to be used as a screening method at this stage of development. CONCLUSION: the collected dataset of MRI and their markup made it possible to obtain an AI model that allows solving the problem of segmenting a rectal tumor and determining its site. The next stage in the development of AI is to improve its specificity, expand the analyzed parameters, such as the depth of tumor invasion, visualization of metastatic lymph nodes and the status of the resection margin. To further develop the model metric and improve its diagnostic capabilities, we should experiment with training parameters and increase the dataset.