Development of a Machine Learning Model to Evaluate Changes during Radiotherapy in Cervical Cancer Tumor Cells.

Abstract

Radiotherapy is considered the standard treatment for advanced cervical cancer. It is known that irradiation changes the morphology of tumor cells during the treatment, though its clinical significance is unknown. Recently, the usefulness of machine learning using medical images has been reported, as also as pathological images. The purpose of this study is to create a classification model for cervical tumor biopsy treated with radiotherapy, visualized on whole slide images using machine learning, in an attempt to explore the clinical significance of morphology changes during radiotherapy. Cervical cancer patients who underwent radical irradiation with both pre- and mid-treatment biopsy from April 1, 2013, to December 1, 2020, were retrospectively analyzed. Tumor biopsies were Hematoxylin and eosin stained and digitized to whole slide images (WSIs). WSI s were divided into tiles of 224 × 224 pixels and converted into feature vectors using a pre-trained convolutional neural network (densenet121). A training and test dataset were divided 1:1 on a patient basis, and the model was trained to classify "pre-treatment" and "mid-treatment", and its accuracy was evaluated. The probability obtained from the classification model was defined as the radiation change index (RCI), and its color map was projected on the WSI for visualization. Survival analysis was performed to examine the clinical significance of the RCI values. A total of eighty-four patients were analyzed, and the median observation period was 3.2 years. 184 WSIs were obtained, and 2203407 tiles were generated. The classification model was trained using 2500 tiles for each "pre-treatment" and "mid-treatment". The accuracy of the classification model was 70.8 %, with an AUC of 0.77 for the ROC curve. The probability obtained from the classification model was defined as RCI, with 0 being "pre-treatment" and 1 being "mid-treatment", and the mean RCI of "pre-treatment" and "mid-treatment" were 0.38 and 0.61, respectively. When the RCIs were visualized with a color map, the areas considered to be "pre-treatment" were consistent with a viable tumor component, and the areas considered to be "mid-treatment" were consistent with fibrosis supposed to cause by irradiation. To evaluate the clinical significance of RCI, we divide the clinical cohort into two groups with a threshold mean RCI of 0.4 in the pre-treatment biopsy. The disease-free survival was 91 months and 47 months in the RCI<0.4 and RCI≥0.4 groups, respectively. We created a model to classify tissues before and during radiotherapy and successfully visualized it on slide images. The low RCI group before treatment had a better prognosis than the high RCI group, suggesting that the tumor morphologic features obtained by machine learning may be useful for prognosis prediction. Further studies are planned to develop a model that can accurately predict prognosis.