PolyDeep is a research project for better polyp detection and classifiation funded by the Spanish Ministry of Science, Innovation and Universities, through the National Programme for Research Aimed at the Challenges of Society 2017 call (DPI2017-87494-R).
The aim of this call is to promote the generation of scientific knowledge aimed at finding solutions to the problems presented in the challenges of society identified in the Spanish Strategy for Science, Technology and Innovation and in the State Plan for Scientific and Technical Research and Innovation, through quality research, evidenced both by its contribution to the solution of social, economic and technological problems and by the publication of its results in forums with a high scientific and technological impact or the internationalisation of activities.
Colorectal cancer is the most common cancer in Spain, with 41.441 new cases in 2015 and 15.449 reported deaths in 2014. This pathology usually initiates from neoplastic lesions, known as polyps, with different levels of malignancy. The diagnosis and management of polyps is based on endoscopic resection, followed by histological diagnosis. If possible, a sufficiently accurate optical diagnosis would avoid the need of histological analysis of small lesions ("resect and discard"), give recommendations of endoscopic follow-up just after polypectomy, as well as "leave in-situ" those polyps present in the rectum-sigmoid which are not adenomatous. In this sense, the optical diagnosis would reduce costs associated to histological analysis and the endoscopic resection associated risks. The NICE international classification (Narrow band imaging International Colorectal Endoscopic classification) proposes the vascular and glandular pit patterns evaluation with the aim of determining the histology employing highdefinition endoscopes provided with Narrow Band Imaging, with or without magnification, available in almost all hospitals. However, the use of NICE can only be recommended in units where endoscopists have previously followed the required training, overcome the learning curve, evaluated the results and confirmed that the minimum requirements are met. In addition, the available evidence regarding the NICE classification use for invasive adenocarcinoma (NICE 3) and the level of invasion and the advanced serrated lesions is very limited and reduced to expert centers. With this motivation, systems able to predict polyp histology, independently of the center and the observer expertise, are required.
Nowadays, with the rise of Deep Learning techniques, specially in the field of medical image analysis, a new generation of Machine Learning algorithms have appeared, which are able to create prediction models reaching unprecedented accuracy levels. A practical application of these kind of techniques would be the detection followed by classification of colorectal polyps, enabling new Computer-Aided Diagnosis systems development, which are not yet provided by the most important colonoscopy equipment manufacturers.
This project proposes the development of POLYDEEP, a real-time computer-aided diagnosis system of colorectal polyps in colonoscopic images, tackling both detection and further classification, reaching a functional prototype for the clinical setting. We will employ Deep Learning techniques over polyp images available in public datasets, as well as images coming from a new image bank which will be also designed and implemented during this project.
The use of POLYDEEP would improve endoscopist's work, who would be provided with an immediate second opinion, especially useful for novice endoscopists. In addition, we expect that POLYDEEP improves colorectal polyp identification, avoiding resections, histological analyses and unnecessary economic costs, as well as the planning of the most adequate treatment (endoscopic or surgical) in adenocarcinomas.
The main objective of this project is the development of a CAD system for automatic detection and classification of colorrectal polyps in real time.
In order to achieve this objective, three specific objectives are necessary:
Deep Learning requires a high amount of data in order to train detection or classification models capable of deal with images. However, the public datasets of colorrectal polyp images that currently available are still small.
Therefore, as part of this project a new public video and image bank of colorrectal polyps will be developed, both for training the Deep Learning models needed and for helping the community to develop their own models.
The polyp video and image bank developed will be available through the Biobank of the Galicia Sur Health Research Institute (IISGS).
Deep Learning has gained major relevance in the field of computer vision (CV), achieving better results than other classical CV techniques. In addition, its ability to automatically identifying image features has simplified the model creation process, at the expense of requiring big image datasets. In this project we expect to take advantage of DL techniques for both detecting and classifying polyps.
At one hand, by using DL object detection techiniques, such as SSD, RCNN or YOLO, we expect to detect and locate polyps in realtime colonoscopy video.
On the other hand, for polyp classification, state-of-the-art DL classification models will be used as a base to develop our own classifier.
A Computer-Aided Diagnosis (CAD) prototype will be developed as a way to translate the developed DL models to the clinic. This prototype will include a custom application that processes endoscopy video stream in real-time using the DL models to add information that can help the endocopist to identify and classify polyps. This application will be integrated into a module that can be added to a endoscopy tower.