Improved mammography classification using data augmentation and transfer learning

bachelor thesis

Ana Peterka (Author), Zoran Bosnić (Mentor), Evgeny Osipov (Co-mentor)

Abstract

Main goal of this thesis is to improve binary classification of mammograms, which could serve as a second opinion to the radiologists and give patients faster results. For this we studied the benefits of using data augmentation techniques and transfer learning. Training a deep convolutional neural network (DCNN) from scratch is difficult, because it requires large amounts of labeled training data. This is a big problem especially in the medical domain, since datasets are scarce and the data is often imbalanced - there is a higher prevalence of healthy results than pathological findings. This can result in overfitting the model. We try to mitigate this issue by generating novel data. We apply affine transformations to images as well as we generate new images that are produced by conditional infilling GAN. Using transfer learning improves classification and speeds the training process of prediction model. Our results show that we can relatively easy generate new and realistic looking data. With the help of transfer learning we can further improve the classification of benign and malignant mammograms.

Keywords

artificial intelligence;machine learning;deep learning;convolutional neural networks;generative adversarial networks;mammography;cancer detection;computer science;computer and information science;computer science and mathematics;interdisciplinary studies;diploma;

Data

Language:	English
Year of publishing:	2022
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[A. Peterka]
UDC:	004.8:618.19-006(043.2)
COBISS:	101446403
Views:	127
Downloads:	25
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	Slovenian
Secondary title:	Izboljšava klasifikacije mamogramov z generiranjem umetnih podatkov in prenosom učenja
Secondary abstract:	Cilj diplomskega dela je izboljšati binarno klasifikacijo mamogramov, ki bi služila za podajanje drugega mnenja radiologom in nudila hitrejše rezultate bolnikom. V ta namen preučujemo dve tehniki, in sicer tehniko za nadgradnjo učnih podatkov in tehniko za prenos učenja. Globoke konvolucijske mreže potrebujejo za učenje obširne baze podatkov, da lahko primerno nastavijo vrednosti parametrov. Teh pa še posebej primanjkuje v medicinski domeni, obenem pa so podatki tudi zelo neuravnovešeni – prevladuje več zdravih kot obolelih primerov. To poskušamo rešiti z umetnim generiranjem novih podatkov, v našem primeru, mamogramov. Najprej bazo povečamo z raznimi transformacijami slik, nato pa še dodatno z generiranjem novih sintetičnih mamogramov s pomočjo generativne nasprotniške mreže ciGAN. Pri sami klasifikaciji mamogramov si pomagamo s prenosom učenja, kjer gre za prenos vrednosti parametrov iz nevronske mreže. Naši rezultati prikazujejo, da lahko relativno učinkovito generiramo sintetične mamograme z realističnim izgledom. Če jih uporabimo v kombinaciji s prenosom učenja, bistveno izboljšajo klasifikacijo v patološke in zdrave mamograme.
Secondary keywords:	globoko učenje;konvolucijske nevronske mreže;generativne nasprotniške mreže;detekcija raka;interdisciplinarni študij;univerzitetni študij;diplomske naloge;Umetna inteligenca;Strojno učenje;Mamografija;Rak (bolezen);Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS):	Bachelor thesis/paper
Study programme:	1000407
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	V, 36 str.
ID:	14768472