Sildistamata andmete kasutamine rinnavähi ennustamise parendamiseks
Files
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Selles väitekirjas esitatakse sildistamata andmeid kasutav süvaõppe lähenemine rinna infiltratiivse duktaalse kartsinoomi koeregioonide automaatseks klassifitseerimiseks rinnavähi patoloogilistes digipreparaatides. Süvaõppe meetodite tööpõhimõte on sarnane inimajule, mis töötab samuti mitmetel tõlgendustasanditel. Need meetodid on osutunud tulemuslikeks ka väga keerukate probleemide nagu pildiliigituse ja esemetuvastuse lahendamisel, ületades seejuures varasemate lahendusviiside efektiivsust. Süvaõppeks on aga vaja suurt hulka sildistatud andmeid, mida võib olla keeruline saada, eriti veel meditsiinis, kuna nii haiglad kui ka patsiendid ei pruugi olla nõus sedavõrd delikaatset teavet loovutama. Lisaks sellele on masinõppesüsteemide saavutatavate aina paremate tulemuste hinnaks nende süsteemide sisemise keerukuse kasv. Selle sisemise keerukuse tõttu muutub raskemaks ka nende süsteemide töö mõistmine, mistõttu kasutajad ei kipu neid usaldama. Meditsiinilisi diagnoose ei saa järgida pimesi, kuna see võib endaga kaasa tuua patsiendi tervise kahjustamise. Mudeli mõistetavuse tagamine on seega oluline viis süsteemi usaldatavuse tõstmiseks, eriti just masinõppel põhinevate mudelite laialdasel rakendamisel sellistel kriitilise tähtsusega aladel nagu seda on meditsiin. Infiltratiivne duktaalne kartsinoom on üks levinumaid ja ka agressiivsemaid rinnavähi vorme, moodustades peaaegu 80% kõigist juhtumitest. Selle diagnoosimine on patoloogidele väga keerukas ja ajakulukas ülesanne, kuna nõuab võimalike pahaloomuliste kasvajate avastamiseks paljude healoomuliste piirkondade uurimist. Samas on infiltratiivse duktaalse kartsinoomi digipatoloogias täpne piiritlemine vähi agressiivsuse hindamise aspektist ülimalt oluline. Käesolevas uurimuses kasutatakse konvolutsioonilist närvivõrku arendamaks välja infiltratiivse duktaalse kartsinoomi diagnoosimisel rakendatav pooleldi juhitud õppe skeem. Välja pakutud raamistik suurendab esmalt väikest sildistatud andmete hulka generatiivse võistlusliku võrgu loodud sünteetiliste meditsiiniliste kujutistega. Seejärel kasutatakse juba eelnevalt treenitud võrku, et selle suurendatud andmekogumi peal läbi viia kujutuvastus, misjärel sildistamata andmed sildistatakse andmesildistusalgoritmiga. Töötluse tulemusena saadud sildistatud andmeid eelmainitud konvolutsioonilisse närvivõrku sisestades saavutatakse rahuldav tulemus: ROC kõvera alla jääv pindala ja F1 skoor on vastavalt 0.86 ja 0.77. Lisaks sellele võimaldavad välja pakutud mõistetavuse tõstmise tehnikad näha ka meditsiinilistele prognooside otsuse tegemise protsessi seletust, mis omakorda teeb süsteemi usaldamise kasutajatele lihtsamaks. Käesolev uurimus näitab, et konvolutsioonilise närvivõrgu tehtud otsuseid aitab paremini mõista see, kui kasutajatele visualiseeritakse konkreetse juhtumi puhul infiltratiivse duktaalse kartsinoomi positiivse või negatiivse otsuse langetamisel süsteemi jaoks kõige olulisemaks osutunud piirkondi.
The following thesis presents a deep learning (DL) approach for automatic classification of invasive ductal carcinoma (IDC) tissue regions in whole slide images (WSI) of breast cancer (BC) using unlabeled data. DL methods are similar to the way the human brain works across different interpretation levels. These techniques have shown to outperform traditional approaches of the most complex problems such as image classification and object detection. However, DL requires a broad set of labeled data that is difficult to obtain, especially in the medical field as neither the hospitals nor the patients are willing to reveal such sensitive information. Moreover, machine learning (ML) systems are achieving better performance at the cost of becoming increasingly complex. Because of that, they become less interpretable that causes distrust from the users. Model interpretability is a way to enhance trust in a system. It is a very desirable property, especially crucial with the pervasive adoption of ML-based models in the critical domains like the medical field. With medical diagnostics, the predictions cannot be blindly followed as it may result in harm to the patient. IDC is one of the most common and aggressive subtypes of all breast cancers accounting nearly 80% of them. Assessment of the disease is a very time-consuming and challenging task for pathologists, as it involves scanning large swatches of benign regions to identify an area of malignancy. Meanwhile, accurate delineation of IDC in WSI is crucial for the estimation of grading cancer aggressiveness. In the following study, a semi-supervised learning (SSL) scheme is developed using the deep convolutional neural network (CNN) for IDC diagnosis. The proposed framework first augments a small set of labeled data with synthetic medical images, generated by the generative adversarial network (GAN) that is followed by feature extraction using already pre-trained network on the larger dataset and a data labeling algorithm that labels a much broader set of unlabeled data. After feeding the newly labeled set into the proposed CNN model, acceptable performance is achieved: the AUC and the F-measure accounting for 0.86, 0.77, respectively. Moreover, proposed interpretability techniques produce explanations for medical predictions and build trust in the presented CNN. The following study demonstrates that it is possible to enable a better understanding of the CNN decisions by visualizing areas that are the most important for a particular prediction and by finding elements that are the reasons for IDC, Non-IDC decisions made by the network.
The following thesis presents a deep learning (DL) approach for automatic classification of invasive ductal carcinoma (IDC) tissue regions in whole slide images (WSI) of breast cancer (BC) using unlabeled data. DL methods are similar to the way the human brain works across different interpretation levels. These techniques have shown to outperform traditional approaches of the most complex problems such as image classification and object detection. However, DL requires a broad set of labeled data that is difficult to obtain, especially in the medical field as neither the hospitals nor the patients are willing to reveal such sensitive information. Moreover, machine learning (ML) systems are achieving better performance at the cost of becoming increasingly complex. Because of that, they become less interpretable that causes distrust from the users. Model interpretability is a way to enhance trust in a system. It is a very desirable property, especially crucial with the pervasive adoption of ML-based models in the critical domains like the medical field. With medical diagnostics, the predictions cannot be blindly followed as it may result in harm to the patient. IDC is one of the most common and aggressive subtypes of all breast cancers accounting nearly 80% of them. Assessment of the disease is a very time-consuming and challenging task for pathologists, as it involves scanning large swatches of benign regions to identify an area of malignancy. Meanwhile, accurate delineation of IDC in WSI is crucial for the estimation of grading cancer aggressiveness. In the following study, a semi-supervised learning (SSL) scheme is developed using the deep convolutional neural network (CNN) for IDC diagnosis. The proposed framework first augments a small set of labeled data with synthetic medical images, generated by the generative adversarial network (GAN) that is followed by feature extraction using already pre-trained network on the larger dataset and a data labeling algorithm that labels a much broader set of unlabeled data. After feeding the newly labeled set into the proposed CNN model, acceptable performance is achieved: the AUC and the F-measure accounting for 0.86, 0.77, respectively. Moreover, proposed interpretability techniques produce explanations for medical predictions and build trust in the presented CNN. The following study demonstrates that it is possible to enable a better understanding of the CNN decisions by visualizing areas that are the most important for a particular prediction and by finding elements that are the reasons for IDC, Non-IDC decisions made by the network.