[ Paper Summary ] Accurate Classification of Diminutive Colorectal Polyps Using Computer-Aided Analysis
I skipped a lot of materials presented in this paper since, the main idea of this paper is that CNN have done a better job at performing diagnostic when compare to novice endoscopists. (If you wish to know about the full details please read the paper.)
Please note that this post is for my future self to look back and review the materials on this paper with out reading it all over again.
Abstract
Narrow band imaging can be used in real-time prediction of histologic features of colorectal polyps, however expertise in narrow-band imaging is required to make accurate prediction. The authors of this paper trained a deep neural network to perform that task. And many measurements axis were used to measure the performance of the algorithm. (In the test set the DNN-CAD identified neoplastic or hyperplastic polyps with 96.3% sensitivity, 78.1% specificity, a PPV of 89.6%, and a NPV of 91.5%.). The authors of this paper have developed a algorithm in which that can identify neoplastic or hyperplastic colorectal polyps less than 5 mm with high accuracy.
Background and Aims
Colorectal cancer is one of the most common malignancies worldwide and is currently the third leading cause of cancer death in Taiwan. And to avoid unnecessary pathologic evaluation and endoscopic resection accurate diagnosis is important. And Narrow-band imaging allows real-time histologic predictions based on colorectal polyps, however expertise is needed to correctly perform diagnosis. Naturally computer-aided diagnosis systems can be used to help the diagnostic process.
Materials and Methods
In this section, the authors briefly discusses the approval of this study via variety of different organization. And how the images were collected (from Tri-Service General Hospital) as well as the fact that they used tensorflow. Additionally, they removed some images that were blurred as well as used the Inception v3 model (with the final layer modified) and trained the network for 4k epochs. (training performance can be seen below.)
Then the authors describes how the algorithm / experts were tested. (test image contains 96 hyperplastic and 188 neoplastic polyps). 2 experts with more than 5 years of colonoscopy experience and 4 novices with 1 year of colonoscopy experience, all of whom were blinded to histologic data have participated in this experiment.
Results
The details of the test images can be seen above, among the 284 polyps, 56 (19.7%) were located in the cecum and ascending colon, 43 (15.1%) in the transverse colon, 12 (4.2%) in the descending colon, 51 (18.0%) in the sigmoid colon, and 122 (43.0%) in the rectum.
And as seen above, deep neural network’s performance was competitive when even compared to experts. It is interesting to note that 3 of the novice endoscopists failed to meet the PIVI criteria. (The Preservation and Incorporation of Valuable Endoscopic Innovations (PIVI) initiatives reported in the American Society of Gastrointestinal Endoscopy guidelines suggest 17 that a histologic assessment of diminutive polyps (5 mm) should provide a 90% NPV for adenoma detection.)
Final Words
If any errors are found, please email me at jae.duk.seo@gmail.com, if you wish to see the list of all of my writing please view my website here.
Meanwhile follow me on my twitter here, and visit my website, or my Youtube channel for more content. I also implemented Wide Residual Networks, please click here to view the blog post.
Reference
- Chen PJ, e. (2018). Accurate Classification of Diminutive Colorectal Polyps Using Computer-Aided Analysis. — PubMed — NCBI . Ncbi.nlm.nih.gov. Retrieved 31 July 2018, from https://www.ncbi.nlm.nih.gov/pubmed/29042219
