Deep learning methods for segmentation of images of frozen tissue sections

doctoral thesis

• Title (croatian): Metode dubokoga učenja za segmentaciju slika smrznutih patohistoloških rezova
• Author: Dario Sitnik
• Mentor: Ivica Kopriva (mentor)
• Committee member: Ivica Kopriva (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronic Systems and Information Processing); Zagreb
• Granter: Ruđer Bošković Institute; (Division of Electronics); Zagreb
• Defense date and country: 2022, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

It is well known that clinical annotation of histopathology data hinders the creation and testing of computer aided diagnostic algorithms related to histopathology image processing. Also, it is important to have correctly annotated data in order to optimize and analyze the performance of any data-driven technique, particularly supervised classification systems. Therefore, ground truth data collection must be led and supervised by experienced medical experts. However, they require a great amount of time to accomplish this task. In this research, a public dataset, CoCaHis, was created to address the shortage of labeled histopathological images at a pixel-level. The dataset contains 82 histopathological images of hematoxylin-eosin stained frozen sections obtained intraoperatively from 19 patients diagnosed with adenocarcinoma of a colon in a liver. Pixel-wise annotations by seven experts and the majority vote between all of them were also included. Experts were four pathologists, one senior student of medicine and two residents in pathology. Corresponding Fleiss’ kappa score of 0.74 shows a substantial inter-annotator agreement. Furthermore, besides originally stained images, the dataset consists of two stain-normalized versions of two sets of images stain-normalized with respect to two target images selected by the pathologist. For the newly formed CoCaHis dataset, two conventional machine learning algorithms and three deep learning based models were developed to provide baseline results. Specifically, incremental kNN, SVM, DeepLabv3+, U-Net, and U-Net++ were used to perform binary (cancer vs. non-cancer) image segmentation. When combining results from originally stained and stain-normalized image sets, DL models achieve even better results. On an independent test set, deep learning models outperformed SVM and kNN by an average of 14% in terms of micro-averaged balanced accuracy (BACC), 15% in terms of micro-averaged F1 score, and 26% in terms of micro-averaged positive predicted value (PPV). On the other hand, performance scores of different DL models fall within 2%. Finally, U-Net++ architecture with DenseNet201 backbone yielded the highest micro-averaged BACC score of 89.34%, F1 score of 83.67% and PPV of 81.11%. The database and tutorials are publicly available at cocahis.irb.hr.

Abstract (croatian)

Dobro je poznato da je označavanje patohistoloških podataka glavni razlog za sporu izradu i testiranje računalno potpomognutih dijagnostičkih sustava povezanih s obradom slike patohistoloških rezova. Važno je imati ispravno označene podatke kako bi se optimirali i analizirali rezultati algoritama, posebno u nadziranom učenju. Stoga prikupljanje oznaka, tj. labela moraju voditi i nadzirati iskusni medicinski stručnjaci. Međutim, taj proces zahtijeva mnogo vremena. Cilj ovog istraživanja bio je kreirati javno dostupnu bazu podataka, CoCaHis, kako bi se nadomjestio nedostatak označenih slika patohistoloških preparata na razini slikovnih elemenata. Skup podataka sadrži 82 slike sekcija smrznutih patohistoloških preparata, obojenih hematoksilinom i eozinom, uzetih intraoperativno od 19 pacijenata s dijagnozom adenokarcinomom debelog crijeva u jetri. Također, u bazi su sadržane oznake slika sedam stručnjaka. Pripadajuća Fleiss' kappa statistika vrijednosti 0.74 pokazuje da među označivačima postoji znatna razina slaganja. Dodatno, uz slike originalno obojenih uzoraka, u bazi su sadržane i slike normaliziranih boja prema dvije ciljne slike koje je odabrao patolog. Kako bi istraživanje na cijelom skupu podataka, pružilo i osnovne rezultate segmentacije s kojima se buduće studije mogu usporediti, upotrijebljena su dva konvencionalna modela strojnog učenja te tri modela dubokog učenja. Konkretno, naučeni i ispitani su inkrementalni algoritam k najbližih susjeda (kNN), inkrementalni stroj potpornih vektora (SVM), DeepLabv3+, U-Net, i U-Net++. Kada su se kombinirali dijagnostički rezultati sa slika originalno obojenih preparata i oni od slika normaliziranih boja, duboki modeli postigli su bolje rezultate. Na nezavisnom testnom skupu, duboki modeli su bili bolji od SVM-a i kNN-a u prosjeku za 14% u mikro-uprosječenoj balansnoj točnosti, 15% u mikro-uprosječenom F1, te 26% u slučaju mikro-uprosječene preciznosti (PPV). S druge strane, uspješnost različitih modela dubokog učenja po tim metrikama je u prosjeku unutar 2% razlike. Konačno, U-Net++ arhitektura s DenseNet201 osnovom je najbolje segmentirala slikovne elemente. Ostvareni rezultati su redom: 89.34% za balansnu točnost, 83.67% za F1 te PPV od 81.11%. Baza podataka i vodič za učitavanje i manipulaciju podatcima su javno dostupni na cocahis.irb.hr.

• Keywords: intraoperativna dijagnostika; adenokarcinom debelog crijeva u jetri; smrznuta sekcija; svjetlosna mikroskopija; neobojeni preparati; strojno učenje; duboko učenje; učenje s prijenosom znanja; segmentacija; hematoxylin-eosin; normalizacija boja.
• Keywords (english): intraoperative diagnosis; adenocarcinoma of a colon in a liver; frozen sections; light microscopy; unstained specimens; deep learning; machine learning; transfer learning; segmentation; hematoxylin-eosin; stain-normalization
• Language: english
• URN:NBN: urn:nbn:hr:168:827717
• Promotion: 2022
• Project: Number: IP-2016-06-5235; Title: Strukturne dekompozicije empirijskih podataka za računalno potpomognutu dijagnostiku bolesti; Acronym: DEDAD; Leader: Ivica Kopriva; Jurisdiction: Croatia; Funder: HRZZ; Funding stream: IP
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Catalog URL: http://lib.fer.hr/cgi-bin/koha/opac-detail.pl?biblionumber=53362
• Type of resource: Text
• Extent: xiii, 94 str.
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2022-12-06 09:46:41