Differentiation of normal and cancerous urothelial cells from microscopic images using machine learning

Mnogi raziskovalci in zdravniki zaradi pomanjkanja dobrih in zanesljivih orodij, mikroskopske slike s celicami še vedno označujejo ročno, kar je časovno potratno in zvišuje stroške raziskovanja in zdravljenja. Kot odgovor na ta problem, smo razvili orodje, ki z metodo Watershed samodejno zaznava in razločuje normalne in rakave urotelijske celice. Orodje v prvih korakih segmentira mikroskopske slike in označi regije odkritih celic. Na osnovi odkritih celic sledi izbor in izračun značilk, ki jih orodje v naslednjih korakih uporabi za učenje napovednih modelov. V raziskavi smo celice v ciljna razreda uvrščali z nevronskimi mrežami, naključnimi gozdovi, naivnim Bayesovim klasifikatorjem, odločitvenimi pravili CN2, metodo podpornih vektorjev ter metodama boosting in bagging. Opisan postopek smo izvedli s samodejno označenimi, nato pa še z ročno označenimi slikami normalnih prašičjih in rakavih humanih urotelijskih celic.Empirična opazovanja kažejo, da orodje dobro segmentira celice. Kljub temu se izkaže, da napovedni modeli boljše razločujejo med normalnimi in rakavimi celicami na ročno označenih celicah. Najboljše rezultate z ročno označenimi celicami dosegajo nevronske mreže (AUC (area under the curve) 0,9052), metoda bagging (AUC 0,9041) in naključni gozdovi (AUC 0,9005). Zmogljivost orodja smo preverili še z naborom citopatoloških urinskih vzorcev. Pri teh vzorcih so rezultati samodejne segmentacije opazno slabši kot pri drugih naborih slik. Kljub temu, bi lahko z nadaljnjimi izboljšavami orodje bistveno pripomoglo k poenostavljenim in zanesljivejšim analizam mikroskopskih slik rakavih celic.As a result of a lack of reliable tooling, much of the cell detection in microscopic imaging is still done manually. This in turn raises research and treatment costs. To tackle this problem, we developed a tool, which automatically detects and classifies normal and cancerous urothelial cells. In the first part the tool segments microscopic images and marks the discovered cell regions. On the basis of the discovered regions, the tool extracts a set of features, which are later used for learning classification models. Neural nets, random forests, naive Bayes classificator, decision rules, SVM, boosting and bagging were used for classification. We used both automatically and manually marked images of normal pig cells and cancerous human cells. Empirical observation shows, that the tool segments cells really well, nonetheless, we noticed that classificators perform better on manually marked cells.The best results were achieved (using manually marked cells) by neural nets (AUC (area under the curve) 0,9052), bagging (AUC 0,9041) and random forests (AUC 0,9005). The performance of the tool was further tested with cytopathological urine samples. The results of image segmentation with these samples were noticeably worse than with other image sets. With future enhancements this tool could considerably contribute to simpler and more reliable microscopic image analysis of cancerous cells