Kundan S Chufal, Irfan Ahmad, Ram Bajpai, Alexis Andrew Miller, Rahul Chowdhary, Munish Gairola
Nov 1, 2020
The purpose of this research paper was to develop an Artificial Neural Network (ANN) model to predict the pathological response after neoadjuvant chemoradiation (NACRT) in patients with esophageal carcinoma, based on Radiomics data extracted from pre-NACRT Computed Tomography (CT) datasets. The study included 97 patients who underwent NACRT followed by radical esophagectomy, with 55 patients achieving a pathological complete response (pCR).
Radiomics feature extraction was performed on the pre-NACRT CT datasets, and Random Forest (RF) classifier was used for feature selection. Multivariable logistic regression further reduced the dimensionality of the selected features. The ANN model, using Multi-Layer Perceptron (MLP) network, was trained and validated on the cohort, with the pathological outcome (pCR present/absent) as the output layer. The model's predictive accuracy was evaluated using the Area Under Curve (AUC) analysis.
A total of 254 features were extracted, and RF identified 15 features with a high probability of predicting the pathological outcome. After logistic regression, 7 features were selected as input for the MLP model. These features described the sphericity and 3-dimensional higher-order characteristics of the tumor. The overall accuracy of the model was 80% in the training cohort and 77.8% in the validation cohort, with an AUC of 0.87.
The study concluded that ANN-based predictive modeling using Radiomics features is feasible for predicting the pathological outcome after NACRT in esophageal carcinomas. Further investigation is warranted to validate and refine the model.
