Melanoma Diagnosis and Localization from Whole Slide Images using Convolutional Neural Networks
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3015481Utgivelsesdato
2022Metadata
Vis full innførselSamlinger
- Studentoppgaver (TN-IDE) [823]
Sammendrag
During the last decade, no other cancer type in Norway have had higherincrease in incidents than skin cancer. Melanoma is the most aggressive typeof skin cancer because it has the ability to rapidly spread, which makes earlydiagnosis important. By conventional methods, melanoma diagnosis involvesexamination of melanocytic lesions under a light microscope performed by apathologist. The diagnosis is both time and labor intensive, subjective andnot always easy to reproduce. The increase in incidents means even moreworkload on the pathology departments, and increased waiting time for thepatients.Digital pathology is a subfield of pathology that has emerged due to theability to scan histology glass slides into digital whole slide images (WSI).This opens for computational analysis typically using image processing anddeep learning. In this thesis a method is proposed to 1) automatically classifyWSIs as either melanoma or benign nevus using image processing andconvolutional neural network, and 2) localize the lesions and display them inoverview images to aid the pathologist and add confidence to 1).The dataset consists of 93 WSIs from Stavanger University Hospital. Allhave been provided slide-based diagnosis labels, and regions of interest havebeen annotated by a pathologist. Due to the giga-pixel size of the WSIs theyhad to be split into patches, and the patch-based labels correspond to theannotations.A pre-trained VGG16 network was fine-tuned on 215320 patches frombenign lesions, 215320 from malignant lesions, and 16806 from other normaltissue types combined into one class. The proposed inference pipeline isto predict all patches of an unseen WSI based on the trained VGG16 andthresholding, create an overview image with the predictions, and providea slide-based prediction based on the ratio of malignant- to benign lesionpixels.The 5 benign nevus and 4 melanoma WSIs of the unseen test set were allpredicted with the correct slide-based label. The overview images correctlylocalized 83.78 and 93.54 % of the benign and malignant lesions, respectively.