Classification of Non-Small Cell Lung Cancer Stage using a Convolutional Neural Network

Gade, Josefine Dam ; Hald, Line Sofie

4. term

Biomedical Engineering and Informatics, Master

2019

2019-06-05

80 pages

Lungekræft er den ledende årsag til kræftrelateret dødsfald, hvoraf ikke-småcellet lungkræft er den hyppigste kræfttype. Denne type er ydermere inddelt i fire undertyper, hvoraf den primære tumor af disse undertyper kan inddeles i fire stadier af kræft. Denne detektering bliver lavet af patologer ud fra histopatologiske billeder. Grundet den mulige variation i antallet af celler i en histopatologisk vævsprøve samt variation i staining intensiteten, er der mulighed for at interobservatør variabilitet opstår ved diagnosticering af tumor undertyper og stadier. Eksisterende metoder har klassificeret undertyper af lungekræft ved brug af CNNs, men det har endnu ikke være undersøgt hvorvidt et CNN kan anvendes til at bestemme stadiet af lungekræft, hvilket vil blive undersøgt i dette studie. Data fra The Cancer Genome Atlas er blevet hentet fra bronchus og lunge sættet. Dataudvælgelse og præprocessering var udført for at forberede dataet til netværket og sikre homogenitet i datasættene. Herefter var AlexNet fine-tunet i de sidste to lag ved brug af initierende vægte fra en prætrænet model. Fire forskellige eksperimenter var udført grundet en bekymring om støjfyldte labels. Derudover var der udført tre WSI-baserede testmetoder for hvert eksperiment til at evaluere de forudsete patch-baserede labels. De trænede modeller var testet patch-baseret. Patchene var klassificerede med en nøjagtighed på 0.52+/-0.04 under de fire eksperimenter. Eksperimenterne med de bedste testresultater havde en nøjagtighed på 0.56 i alle tre WSI-baserede klassificeringsmetoder. Ydermere havde to af de WSI-baserede klassificeringsmetoder den højeste nøjagtighed på 0.56+/-0.01.

Lung cancer is the leading cause of cancer-related mortality, of which non-small cell lung cancer is the most common type, that furthermore is divided into four subtypes. The primary tumor of these subtypes can be divided into four stages of cancer. This is performed by pathologists on histopathological images. However, due to the number of cells in a sample and the staining intensities which can vary, inter-observer variability is likely to occur when diagnosing tumor subtypes and stages. Existing methods have classified lung cancer into subtypes using CNNs, but it has not yet been investigated whether a CNN could be used to determine the stage of lung cancer, which will be investigated in this study. Data from The Cancer Genome Atlas has been acquired from the bronchus and lung subset. Data selection and preprocessing were performed to prepare the data and ensure homogeneity in the datasets, and the AlexNet was fine-tuned using a pretrained model. Four different experiments were conducted due to a concern about noisy labels. Furthermore, three WSI-classification approaches to evaluate the predicted labels were performed in each of the experiments. The trained model was tested patch-based and WSI-based. The patches were classified with an accuracy of 0.52+/-0.04 during the four experiments. The experiments with the best results had an accuracy of 0.56 in all three approaches. Additionally, two of the approaches had the highest accuracies of 0.56+/-0.01 when classifying WSI-based.

Download
View record in AAU Student Projects