The objective has been to compare success of the texture classifier and a human observer in diagnosis of the autoimmune thyroiditis from B-mode ultrasound images and to determine inter- and intra-observer variability. The data set of 161 subjects was classified by four human observers and by the Bayes classifier based on the texture features to three classes (healthy, border state, autoimmune thyroiditis). Two observers had a higher success rate when classifying the healthy class (74.4% and 83.3%), the other two observers classified better cases with autoimmune thyroiditis (59.0% and 77.4%). The classifier gave the relatively high and balanced success rate for both classes (100,0% for healthy and 87.5% for thyroiditis). The different observers' success rates resulted in the high inter- observer variability, showing only a fair agreement among the human observers. There was no significant difference among human observers in the intra-observer variability. Due to the fair agreement among observers in the diagnosis of autoimmune thyroiditis from ultrasound images and good results of the classifier, the best way in establishing diagnosis is computer- aided diagnosis combined with observers' clinical experience.