Data-driven learning approach used for segmentation of histological images

Background. Microscopic colitis (MC) is an inflammation that affects the inner lining of the colon. This could lead to chronic and watery diarrhea, abdominal pains, faecal incontinence, and weight loss. The symptoms of the disease affect overall quality of life for the patient. Diagnosis of MC is done via biopsies examination by pathologist who checks the tissue under a microscope and looks for signs of the disease (increased inflammatory infiltrate in the lamina propria without significant crypt architectural distortion and intraepithelial lymphocytosis). The diagnosis is subjective and could lead to diagnostic discrepancies between evaluators and treatment plan failures. This study focuses on the development of an algorithm for pathologist to assist them in MC diagnosis. Aim. The aim of the current study was to develop an algorithm for MC localisation from microscopic images. Methods. Six hundred histological microscopic images (10 patients) were used to train the elaborated algorithm. Image pre-processing was done by dividing each image into small image elements (superpixels), using simple linear iterative clustering algorithm. After this procedure we applied machine learning approach to classify the pre-processed elements to disease and healthy tissues groups. Machine learning algorithm was pretrained using images, annotated by three pathologists. The proposed algorithm together with image pre-processing was implemented in MatLab development environment. Results. The model was tested on additional microscopic images from five patients (300 images). It showed accuracy of correct classification of analysed microscopic images equal to 0.81, sensitivity — 0.8 and specificity — 0.81. Conclusion. The proposed algorithm can assist pathologists in MC diagnosis and make it more objective.