Isabelle Stålberg and Linnéa Törnblom presents their master's thesis "Development of image analysis for classification of lymphoma from fine needle aspirations"

Examiner: Niels Christian Overgaard

Supervisors: Kalle Åström, Mats Ehinger, Erik Wistén

Abstract:

The majority of lymphoma diagnostics include invasive, surgical procedures which is both unpleasant for the patient, and expensive as well as time consuming for the health-care sector. Due to this, the main purpose of this thesis was to investigate how a previously created image analysis tool could be further developed using a larger dataset with more diagnosis types, as well as a more robust and automated preprocessing and evaluation protocol. Another purpose was to explore whether the algorithm could reach any level of explainability. Using MATLAB, a preprocessing protocol was implemented to create two sets of images of different sizes from the original DICOM images. These sets were then used to train Convolutional Neural Networks in three constellations of classes; benign and malignant; benign, small cell lymphoma, and large cell lymphoma; as well as benign and each individual diagnosis. Using the created evaluation protocol, it was determined that while classification accuracy was considerably decreased from the original image analysis tool, classification between two classes could be achieved with an accuracy of 75.48% using the image set of size 400x400 pixels cropped from original images with lower resolution. The image set of size 200x200 pixels cropped from original images with higher resolution did, however, perform better on entire patients than on individual images, implying that the combination of the more detailed images could be ideal for clinical implementation.