Avtomatična detekcija artefaktov v računalniški radiografiji : magistrsko delo

Grom, Karmen (avtor), Žibert, Janez (mentor), Izlakar, Jani (komentor)

Uvod: Dandanes splošna radiologija še vedno predstavlja primarno slikovno tehniko, zato je potreba po izboljšanju kakovosti rentgenskih slik več kot utemeljena. Kljub velikemu tehnološkemu napredku digitalne radiografije pa se v praksi še vedno srečujemo z artefakti, ki vodijo k slabi kakovosti radiogramov. Namen: Namen je avtomatizirati zaznavanje artefaktov na CR sistemu s pomočjo konvolucijske nevronske mreže. Metode dela: Po vnaprej določenem protokolu smo pridobili vzorec 269 homogenih slik, pridobljenih na 31 CR ploščah. Celotni vzorec je bil ocenjen s strani dveh ocenjevalcev, ki sta določala prisotnost artefaktov. Artefakte smo razvrstili v pet skupin - prah, umazanija in razpoke, ghosting in nehomogenost, ravne linije in proge, nazobčani artefakti in drugo. Slike smo obdelali ter vnesli v prilagojeno in doučeno nevronsko mrežo AlexNet. Za učenje razpoznavanja posamezne skupine artefaktov smo zbrani vzorec razdelili v učni (80 %) in testni del (20 %). Model nevronske mreže za vsako testirano sliko poda verjetnost za prisotnost artefakta. Zaključni korak je predstavljala validacija naučenih modelov nevronskih mrež za vsak artefakt posebej, kjer smo za glavno mero učinkovitosti izbrali AUC mero. Rezultati: Najboljše rezultate smo dosegli pri skupinah nazobčani artefakti (AUC=100 %), drugo (AUC=99,02 %) in ghosting ter nehomogenost (AUC=97,62 %). V skupini nazobčanih artefaktov in drugo smo zajeli malo predstavnikov omenjenih artefaktov, vendar kljub temu lahko ugotovimo, da smo uspešno zaznali obe skupini, verjetno tudi zaradi izstopanja artefaktov. Slabšo detekcijo artefaktov pa lahko opazimo pri skupinah lokalnih artefaktov - prah, umazanija in razpoke (AUC=83,70 %) ter linije in proge (AUC=81,21 %), saj predpriprava slike za vnos v nevronsko mrežo AlexNet, zahteva skaliranje slike na dimenzije 227×227 pikslov. Artefakti manjših dimenzij se posledično s skaliranjem slike izgubijo. Razprava in zaključek: Metoda globokega učenja z doučevanjem in prilagoditvijo konvolucijske nevronske mreže AlexNet, se je izkazala za zelo učinkovit model za detekcijo artefaktov na CR slikah. Iz rezultatov analize lahko zaključimo, da je avtomatična detekcija artefaktov s pomočjo nevronske mreže učinkovita predvsem pri zaznavanju globalnih artefaktov. Kljub vsemu pa lahko ugotovimo visoko zmožnost zaznavanja najpogosteje prisotnih artefaktov in s tem posledično tudi potencialno uporabo sistema za avtomatično detekcijo artefaktov v kliničnem okolju.

• Jezik: Slovenski jezik
• Ključne besede: kontrola kakovosti, računalniška radiografija, artefakti, umetna inteligenca, konvolucijska nevronska mreža
• Vrsta gradiva: Magistrsko delo/naloga
• Organizacija: ZF - Zdravstvena fakulteta
• Leto izida: 2019
• PID: 20.500.12556/RUL-107510
• COBISS.SI-ID: 5611883
• Datum objave v RUL: 21.04.2019

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Automatic detection of artefacts in computed radiography : master thesis

Izvleček

Introduction: Nowadays, general radiography still represents primary imaging technique, therefore the need to improve the quality of x-ray images is more than justified. Despite the great technological advancement of digital radiography, we still detect artefacts, which lead to poor quality radiograms. Purpose: Purpose of study was to develop automated detection of artefacts on CR images with convolutional neural network (CNN). Methods: According to a predetermined protocol, we obtained 269 homogeneous images with 31 CR plates. Colection was evaluated by two experts who determined the presence of artifacts. Artifacts were classified into five groups - dust, dirt and cracks, ghosting and non-uniformity, straight lines, serrated artefacts and others. The images were preprocessed and entered into CNN AlexNet, which was custumized with transfer learning process. In order to learn the recognition of each group of artifacts, we divided colection of images into learning (80 %) and testing (20 %) part. The CNN gives the probability of the presence of an artefact for each image. Validation of learned neural network for each group of artifacts was the final step, where the AUC measure was chosen as the main measure of effectiveness. Results: The best results were achieved in groups of serrated artefacts (AUC= 100 %), others (AUC= 99.02 %) and ghosting and non-uniformity (AUC=97.62 %). In the group of serrated artifacts and others we captured only a few representatives of these artefacts, however, we can conclude that both groups were successfully detected, probably because of the easy noticeable look of artefact. Slightly lover detection of artefacts can be observed in groups of local artifacts - dust, dirt and crack (AUC= 83.70 %), and lines (AUC= 81.21 %), since the preprocessing of an image for input into the CNN AlexNet requires scaling the image dimensions to 227×227 pixels. Artifacts of smaller dimensions are therefore lost by scaling the image. Discussion and conclusion: Deep learning method by teaching and adapting the CNN AlexNet has proven to be very efficient model for detection of artefacts in CR images. From the results of the study we can conclude that the automatic detection of artefacts with the CNN is most effective in detecting global artefacts. However we can confirm high ability of system to detect the most frequently present artefacts and, consequently, the potential use of the system in a clinical environment.

• Ključne besede: quality control, computer radiography, artefacts, artificial intelligence, convolutional neural network