Author: Beth Bradshaw Ghavidel, Benyamin Khajetash, Yang Lei, Meysam Tavakoli 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, Department of Medical Physics, School of Medicine, Iran University of Medical Sciences, Emory University, Department of Radiation Oncology, Emory University 🌍
Purpose: Pancreatic cancer is among the most aggressive types of cancer, with a five-year survival rate of approximately 10%. Recent studies show that radiomics and deep learning (DL)-based methods are promising imaging-based strategies for early and accurate diagnosing of pancreatic cancer and improving patient outcomes. This study aims to combine radiomics and DL methods to develop a computer-aided diagnosis (CAD) scheme of computed tomography (CT) images for automatic diagnosis of pancreatic.
Methods: CT data of 220 patients were collected for this study. We present a hybrid model combining a radiomics model with a 3D convolutional neural network (3D-CNN) that leverages spatial contextual information. The output probabilities of these models are fused for final prediction. The dataset includes CT images of 220 patients (100 with pancreatic cancer and 120 without). Radiomics features, including intensity, texture, and geometric attributes, were extracted from contoured tumors in the CT images, while the 3D-CNN model explored tumor structures. An evidential reasoning (ER) algorithm combined the output probabilities from the two models. Performance was evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC).
Results: Five-fold cross-validation revealed that the hybrid model combining radiomics and 3D-CNN achieved the highest performance. The fused model performed an average AUC of 0.80.
Conclusion: This study demonstrates the potential of integrating radiomics and DL for reliable diagnosis of pancreatic cancer, facilitating improved treatment planning. This study also laid the groundwork for advancing and refining CAD systems by incorporating more sophisticated image processing and machine learning techniques, aiming to achieve more accurate and reliable detection and classification of pancreatic tumors in the future.