Detecting Gastritis from Histological Sections using Machine Learning

Abstract

This work is about detecting gastric stomach glands in histological images. This was done using supervised machine learning and image recognition techniques. The glands were identified and located using image segmentation, on which basis the glands were identified, using classification algorithms. The data was provided by a hospital with histoligical images labeled by medical professionals. For the segmentation, the Cellpose algorithm was used as it yields great results out of the box and provides additional finetuning parameters. For the object detection, a Mask R-CNN model was constructed. Due to multiple possible factors, such as the large amount of required data points and a poor labeling quality, the Mask R-CNN model yielded rather poor results. For a different approach, a numerical feature vector was extracted from each gland segment, containing multiple numerical values, such as the Region Size, Mean Intensity or Circularity of each gland. Then multiple classifiers, such as Random Forests, K-Nearest Neighbours, and Support Vector Machines were used in order to classify the stomach glands based on their feature vectors alone. The best results yielded the Random Forest Classifier, achieving an accuracy of 82.48% on the test set.

Date of Award	2024
Original language	English (American)
Supervisor	Stephan Winkler (Supervisor)