An artificial intelligence (AI) model for automated classification of colorectal polyps could benefit cancer screening programs by improving efficiency, reproducibility, and accuracy, as well as reducing access barriers to pathological services, according to a study by researchers from Dartmouth's and Dartmouth-Hitchcock's Norris Cotton Cancer Center.
A computer science and clinical research team led by Dr Saeed Hassanpour, trained a deep neural network to distinguish the four major types of colorectal polyps at the level of practicing pathologists, as evaluated on a dataset across multiple external institutions. The model also proved that a model designed using data from a single institution can achieve high accuracy on outside data.
"Our study is one of the first to show a deep neural network that is generalizable to data from multiple external medical centres," said Hassanpour. "A challenge in the field of deep learning for medical image analysis is collecting widespread data. Here, we have access to histopathology slides from 24 different institutions, which gave us the opportunity to evaluate and show that the AI models that we train are broadly generalizable to new data from outside."
The deep neural network still performed with the same level of sensitivity and accuracy as practicing pathologists when used on 238 slides spanning 24 different institutions in the US. The study, ‘Evaluation of a Deep Neural Network for Automated Classification of Colorectal Polyps on Histopathology Slides,’ was published in JAMA Network Open.
The access to a multi-institutional dataset was made possible by Hassanpour's collaboration with Dr. Dr Arief Suriawinata, and his group from the Department of Pathology & Laboratory Medicine at Dartmouth-Hitchcock Medical Center and Dr Elizabeth Barry from the Department of Epidemiology at the Geisel School of Medicine at Dartmouth, as well as her colleagues from the Vitamin D/Calcium Polyp Prevention Clinical Trial.
This prognostic study used histopathologic slides collected from January 2016 to June 2016, from Dartmouth-Hitchcock Medical Center, Lebanon, New Hampshire, with 326 slides used for training, 157 slides for an internal data set and 25 for a validation set. For the external data set, 238 slides for 179 distinct patients were obtained from 24 institutions across 13 US states.
The accuracy, sensitivity and specificity of the model to classify four major colorectal polyp types: tubular adenoma, tubulovillous or villous adenoma, hyperplastic polyp, and sessile serrated adenoma, was assessed. Performance was compared with that of local pathologists’ at the point of care identified from corresponding pathology laboratories.
For the internal evaluation on the 157 slides with ground truth labels from five pathologists, the deep neural network had a mean accuracy of 93.5% (95% CI, 89.6%-97.4%) vs local pathologists’ accuracy of 91.4% (95% CI, 87.0%-95.8%). On the external test set of 238 slides with ground truth labels from five pathologists, the deep neural network achieved an accuracy of 87.0% (95% CI, 82.7%-91.3%), which was comparable with local pathologists’ accuracy of 86.6% (95% CI, 82.3%-90.9%).
“In this study, the performance of the deep learning model was similar to that of local pathologists on the internal and external test sets,” the authors concluded. “If confirmed in clinical trials, this model could improve the efficiency, reproducibility, and accuracy of colonoscopy.”
Hassanpour's team has built a graphical user interface for showing the classifications of the neural network. They are currently working on a clinical trial to evaluate the use of their algorithm for assisting pathologists in diagnosis of colorectal polyps.
“We are hoping to create a software application that can help pathologists improve their accuracy, efficiency, and consistency in diagnosing slides," he added.
To access this paper, please click here