Machine learning using distance-based similarity labels

Abstract

The invention relates to a computer-implemented self-supervised learning method for digital pathology. The method comprises: Receiving (102) a plurality of digital images each depicting a tissue sample; Splitting (104) each of the received images into a plurality of tiles; automatically generating (106) tile pairs (312, 313, 314, 315, 422), each tile pair having assigned a label (403) being indicative of the degree of similarity of two tissue patterns depicted in the two tiles of the pair, wherein the degree of similarity is computed as a function of the spatial proximity (d1, d2, d3, d4) of the two tiles in the pair, wherein the distance positively correlates with dis- similarity; and training a machine learning module - MLM - (400, 402, 403, 600) using the labeled tile pairs as training data to generate a trained MLM, the trained MLM being adapted for performing an image analysis of digital histopathology im- ages.

Description

technical field [0001] The invention relates to the field of digital pathology, in particular to the field of image analysis. Background technique [0002] Computational analysis of digital tissue images in the field of digital pathology has a wide range of important biomedical applications, such as tissue detection, segmentation, morphometry, identification and classification of diseases (eg, cancer) and possible treatment options. Currently, machine learning methods are used to address the complexity and variety of image analysis problems. [0003] Typically, supervised machine learning methods are used to solve image analysis problems. Therefore, a machine learning module (MLM) is trained on a set of training images labeled as ground truth by domain experts (especially pathologists and clinicians). During the training phase, the statistical model of the MLM learns to map the relevant image features computed by the image analysis algorithm to the labels contained in the ...

AI Technical Summary

Problems solved by technology

Therefore, the predictive accuracy of MLMs trained on such inconsistent training datasets may be degraded by a large part of "annotation inconsistency / annotation noise"

[0008] For the above reasons, the lack of annotated training datasets of sufficient size and quality is the main reason why many open biomedical problems cannot be addressed and solved by machine learning algorithms already available today

Images