PhD Scientific Days 2020

Pathology and Oncology I. Posters

Artificial Intelligence Based Colorectal Cancer Screening Decision Support

Előadó neve

Pesti, Adrián, MSc

Előadó munkahelye

2nd Department of Pathology, Semmelweis University

Előadó telefonszáma

+36305950409

Előadó e-mail címe

adrianpesti@gmail.com

Az előadás címe

Artificial Intelligence Based Colorectal Cancer Screening Decision Support

Szerző(k) neve és munkahelye

A. Pesti1, E. Kontsek1, B. Pataki2, A. Olar2, D. Ribli2, P. Pollner3, I. Csabai2, A. Kiss1
1 2nd Department of Pathology, Semmelweis University, Budapest
2 Department of Physics of Complex Systems, Eötvös Loránd University, Budapest
3 MTA-ELTE Statistical and Biological Physics Research Group of the Hungarian Academy of Sciences, Budapest

Szekció

Pathology and Oncology I. Posters

Language of the presentation

English

Section, first choice

Pathology and Oncology

Section, second choice

Health Sciences

Összefoglaló szövege

Introduction
Hungary leads the colorectal cancer morbidity and mortality statistics worldwide. To reduce the prevalence of this disease a nationwide screening program has been started in 2018. Patients are checked by colonoscopy after producing positive stool test. The elevated number of colorectal biopsies might be prefiltered by an Artificial Intelligence (AI) algorithm.
Aims
The aim is to help pathologists speed up the process of the true negative cases filtration, and AI make diagnostic annotations on the positive slides. As a result the pathologists have more time to devote on the complex cases, and to form improved diagnoses.
Method
712 HE stained colorectal biopsy slides were gathered from the archive of the 2nd Department of Pathology Semmelweis University. The slides were scanned via P1000 digital slide scanner (3DHistech Ltd.). Compressed, lossy and lossless datasets were formed after the scanning process. A Convolutional Neural Network (CNN) was trained on 612 slides and 100 slides were used as a testing dataset. The pictures have been pre-processed by dropping the background and cutting the images into 512×512 pixel image patches.
All whole slides were annotated by resident doctors and every single annotation was supervised and validated by board certified pathologists. The annotation process involved: 1) global, textual annotation for general diagnosis, 2) local, textual annotation for tagging specific tissue parts, 3) graphical, pixel level annotation for denoting the area of locally annotated tissue parts.
Results
The performance of the CNN was measured with the Area Under receiver operating characteristic Curve (AUC) score for the local labels. The CNN was least successful for the „suspicious invasion” category with 83% AUC while the best result was achieved for the „tumour necrosis” where the learner reached higher than 99% AUC score. Note, that this is an ongoing project, hence we expect further improvement.
Conclusion
Nowadays the legal environment does not allow to produce only algorithm based medical report without human supervision, signature and accountability. However, the workload on the pathologists might be reduced by a built-in decision support module into the digital pathology software infrastructure, resulting doctors to focus on the complicated cases, and supervise the AI.

Additional Information

András Kiss
kiss.andras@med.semmelweis-univ.hu

Acknowledgement: This work was partially supported by NKFIH-837-6/2019, NKFIH-128881/2018 and NKFIH TKP/2019

Bemutatás módja

Poszter

Kind

Szabad

Status

elfogadva

Accepted presentation method

poszter

Előadás fájl jóváhagyás

nem rendelkezett róla

Előadó

4756

Start

11:32

End

11:35

Authors (legacy)

A. Pesti1, E. Kontsek1, B. Pataki2, A. Olar2, D. Ribli2, P. Pollner3, I. Csabai2, A. Kiss1
1 2nd Department of Pathology, Semmelweis University, Budapest
2 Department of Physics of Complex Systems, Eötvös Loránd University, Budapest
3 MTA-ELTE Statistical and Biological Physics Research Group of the Hungarian Academy of Sciences, Budapest