Deep Neural Network for the Prediction of KRAS Genotype in Rectal Cancer
Waleed M Ghareeb1,2, Eman Draz2,3,4, Khaled Madbouly5
1From the Gastrointestinal Surgery Unit (Ghareeb, Hussein), Faculty of Medicine, Suez Canal University Hospitals, Ismaila, Egypt.
Journal of the American College of Surgeons
|August 16, 2022
Summary
This study developed a deep neural network (DNN) to predict KRAS genotype from histopathology images, improving colorectal cancer treatment planning. The DNN model demonstrated high accuracy, potentially reducing costs and saving time for patient screening.
Area of Science:
- Computational pathology
- Oncology
- Genomics
Background:
- KRAS mutation status is critical for colorectal cancer treatment but often remains unchecked due to high costs.
- Hematoxylin and eosin (H&E)-stained histopathological images offer a potential resource for KRAS genotype prediction.
Purpose of the Study:
- To develop and validate a deep neural network (DNN) for predicting KRAS genotype directly from H&E-stained histopathological images of colorectal cancer.
- To assess the performance of the DNN models against human pathologists and evaluate potential cost savings.
Main Methods:
- Three DNN models (KRAS_Mob, KRAS_Shuff, KRAS_Ince) were developed using MobileNet, ShuffleNet, and Inception backbones.
- Large datasets from The Cancer Genome Atlas (49,684 tiles) and an independent cohort (43,032 tiles) were used for training, internal, and external validation.
- Model performance was evaluated using area under the receiver operating curve (AUC) and compared to pathologist performance.
Main Results:
- The KRAS_Mob model achieved the highest AUC (0.8), outperforming other DNNs and surpassing the accuracy of two independent pathologists (AUC 0.79 vs. 0.51).
- A combined prediction approach using KRAS_Mob and KRAS_Shuff showed improved performance.
- The DNN models demonstrated robust predictive capabilities on external validation data.
Conclusions:
- DNNs can accurately predict KRAS genotype from H&E-stained histopathological images, offering a valuable tool for colorectal cancer management.
- This algorithmic approach can serve as a screening method to prioritize patients for laboratory confirmation, potentially leading to significant time and economic savings.
