Artifacts removal from tissue images

Abstract

The invention relates to a digital pathology method of generating program logic configured for removing artifacts from digital tissue images. The method comprises generating, for each of a plurality of original images, a first artificially degraded image by applying a first image-artifact-generation logic on each of the original images; and generating the program logic by training an untrained version of a first machine-learning logic that encodes a first artifacts-removal logic on the original images and their respectively generated first degraded images; and returning the trained first machine-learning logic as the program logic or as a component thereof. The first image-artifact-generation logic is A) an image-acquisition-system-specific image-artifact-generation logic or B) a tissue-staining-artifact-generation logic.

Description

technical field [0001] The present invention relates to image analysis, and more particularly to artifact removal in tissue images. Background technique [0002] In digital pathology, information encoded in digital images of tissue slides is extracted from the images to answer various biomedical questions, for example to assist healthcare professionals in the diagnosis and treatment of disease. The field of digital pathology is currently considered one of the most promising avenues in diagnostic medicine for better, faster and cheaper diagnosis, prognosis and prediction of cancer and other important diseases. Digital pathology technologies are also widely used in drug development settings to assist pathologists in understanding the microenvironment of tumors, how patients respond, how drugs work, and other information that can be obtained from tissue images. [0003] Scanned digital tissue images, especially at high magnifications, tend to have several types of noise associ...

AI Technical Summary

Problems solved by technology

[0007] A problem associated with current denoising methods is that the generation of the noise removal logic involves a training phase of the machine learning logic on training data that is difficult to obtain in the required quantity and quality: often, images must be manually annotated for generation A training dataset where noise artifacts and / or tissue structures that look similar to the artifacts but should not be identified as noise artifacts are labeled as "true artifacts" or "real tissue structures"

The process is time consuming and requires hours, days or even weeks of work by one or more experienced pathologists

Furthermore, manual annotation of "true noise" and "real tissue" may be subjective, and different pathologists may have different insights into the nature of specific image parts that cannot be definitively classified as artifacts

Therefore, the creation of the artifact removal program logic currently requires the creation of the training dataset, which is very time consuming and thus expensive

Images