Uporaba algoritma globokega učenja YOLO za detekcijo objektov na mamografskih slikah

BIZILJ, EVA (author), Brodnik, Andrej (mentor), Žibert, Janez (co-mentor)

V diplomski nalogi smo uporabili algoritem globokega učenja YOLO za detekcijo štirih ključnih objektov (mamile, vrha dojke, dna dojke in prsne mišice) na mamografskih slikah. Za učenje modelov smo uporabili podatkovno zbirko 308 mamografskih slik, ki je v lasti Zdravstvene fakultete, Univerze v Ljubljani. Mamografske slike, ki so narejene v projekcijah CC in MLO smo naključno razdelili v učno (~ 70 % slik), validacijsko (~ 15 % slik) in testno množico (~ 15 % slik). Slike smo označili z orodjem za označevanje slik CVAT. Nato smo na različnih velikostih slik izvedli učenje in vrednotenje modelov z algoritmom YOLOv3 in YOLOv5. Vsi modeli so bili prednaučeni na podatkovni zbirki slik MS COCO. Dobljene modele smo med seboj primerjali po uspešnosti detekcije objektov z uporabo srednje povprečne natančnosti (mAP), natančnosti in priklica. Z algoritmom YOLOv5 smo dobili najboljša modela. Model YOLOv5 je na CC slikah dosegel 88,2 % mAP, 91,8 % natančnost in 87,5 % priklic na testni množici, model YOLOv5 na MLO slikah pa 99,5 % mAP, 99,3 % natančnost in 94,4 % priklic na testni množici.

• Language: Slovenian
• Keywords: globoko učenje, detekcija objektov, mamografija, konvolucijska nevronska mreža, algoritem YOLO
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-144200
• COBISS.SI-ID: 142963971
• Publication date in RUL: 03.02.2023

Secondary language

• Language: English

Abstract

In this thesis we have used the YOLO deep learning algorithm for detection of four key objects (mammilla, the top of breast, the bottom of breast, pectoral muscle) on the mammogram images. To train the models, we have used a database of 308 mammogram images, owned by the Faculty of Health Sciences, University of Ljubljana. The mammogram images in the CC and MLO projections were randomly distributed in a training, validation and test data set. We have labelled the images with an image annotation tool CVAT. We have then performed training and evaluation of models with the YOLOv3 and YOLOv5 algorithms on different size images. All models were pretrained on a MS COCO database. Finally, we compared the models according to their performance. Model YOLOv5, which was trained on CC images, achieved a 88,2 % mAP, 91,8 % precision and 87,5 % recall on the test set. Model YOLOv5, which was trained on MLO images, achieved 99,5 % mAP, 99,3 % precision and 94,4 % recall on the test set.

• Keywords: deep learning, object detection, mammography, convolutional neural network, algorithm YOLO