US-12620086-B2 - Domain aware medical image classifier interpretation by counterfactual impact analysis

Abstract

A neural network, trained for the task of deriving the attribution of image regions that significantly influence classification in a tool for pathology classification, comprising (i) a contracting branch, (ii) an attenuation module, (iii) an interconnected upsampling branch, and (iv) a final mapping module.

Inventors

• Dimitrios LENIS
• David Major
• Maria WIMMER
• Gert Sluiter
• Astrid Berg
• Katja Buehler

Assignees

• AGFA HEALTHCARE NV
• VRVIS ZENTRUM FÜR VIRTUAL REALITY UND VISUALISIERUNG FORSCHUNGS-GMBH

Dates

Publication Date

20260505

Application Date

20210602

Priority Date

20200622

Claims (1)

1. 1 . A neural network, trained for the task of deriving the attribution of image regions that significantly influence classification in a tool for pathology classification, in which a medical image is consecutively processed through (i) a contracting branch, (ii) an attenuation module, (iii) an interconnected up-sampling branch, and (iv) a final mapping module, wherein (i) the contracting branch is fed with said medical image, and derives the same features relating to a pathology of interest at different progressively decreasing resolution scales as said classification tool, (ii) the attenuation module is coupled to the contracting branch and comprises convolutional layers, weighs the final output of the contracting branch and thereby selectively damps the forwarding of a subgroup of said features relating to a pathology of interest, (iii) the up-sampling branch refines this initial localization of the attenuation module, by processing the result of the attenuation module repeatedly through (a) up-sampling, (b) merging with attention-gate weighted feature-maps of the corresponding resolution-scale of the contraction path and (c) convolutional layers, so as to refine the localization of said features of said sub-group, (iv) the final mapping module projects the features onto a two-dimensional grid of same size as said medical image and derives the attribution through a combination of thresholding and smoothing applied to the result of the projection, wherein said neural network is trained by (a) feeding a plethora of learning data, structured in an assortment of batches of at least one single medical image, and (b) assessing said network's derived final output and modifying weights of said neural network such that repeating said assessmentyields an improved output and repeating steps (a) and (b) until no further improvement is obtained, wherein said learning data comprise a plurality of medical images, representative of an anatomical structure of interest, adhering to prerequisites and restrictions of said pathology classification tool, and said assessment of the network's derived final output comprises subsequently marginalizing attributed image regions' contribution towards pathology classification by said classification tool for each image of a batch by altering for each image ofa batch attributed pixels such that they do not affect the outcome of the classification, re-classifying an altered batch using said classification tool and quantifying the result of the re-classification.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This patent application is the U.S. national phase of copending International Patent Application No. PCT/EP2021/064763, filed Jun. 2, 2021, which claims the benefit of European Patent Application No. 20181403.5, filed Jun. 22, 2020. FIELD OF THE INVENTION The present invention is in the field of clinical decision support. More specifically the invention relates to a method for finding and visualizing image regions that significantly influence classification in a tool for pathology classification in a medical image. Background of the Invention The last decade's success of machine learning methods for computer-vision tasks, has driven a surge in computer assisted prediction for medicine and biology. This has posed a conundrum. Current predictors, predominantly artificial neural networks (ANNS), learn a data-driven relationship between input-image and pathological classification, whose validity, i.e. accuracy and specificity, can be quantitatively tested. In contrast, this learned relationship's causality typically remains elusive. A plethora of approaches have been proposed that try to fill this gap by explaining causality through identifying and attributing salient image-regions responsible for a predictor's outcome. Lacking a canonical mapping between an artificial neural network's (ANN's) prediction and its domain, this form of reasoning is predominantly based on local explanations (LE), that is explicit attribution-maps characterising concrete image-prediction tuples. Typically, these attribution-maps are only loosely defined as regions with maximal influence towards the predictor, implying that any texture change within the attributed image-area will significantly change the predictors outcome. Besides technical insight these LE can provide a key benefit for clinical applications: By relating the ANN's algorithmic outcome to the user's a-priori understanding of pathology-causality, they can rise confidence in the predictor's outcome, thereby increase their clinical acceptance. However, some additional restrictions and clarifications are needed to achieve this goal. Qualitatively, such maps need to be informative for its users, i.e. narrow down on regions of medical interest, hence coincide with medical knowledge and expectations. Furthermore, the regions characteristic, i.e. the meaning of maximal influence, must be clearly conveyed. Quantitatively, such LE need to be faithful to the underpinning predictor, i.e. dependent on architecture, parametrization, and preconditions e.g. training-set distribution. The dominant class of methods follow a direct approach. They build upon an ANN's assumed analytic nature, and its layered internal architecture, typically utilizing a modified backpropagation approach to backtrack the ANNs activation back to the input-image. While efficiently applicable, the resulting maps lack a clear a-priori interpretation, are potentially incomplete, coarse, and as shown by recent work, possibly independent of their classifier, hence possibly delivering misleading information. Thereby they are potentially neither informative nor faithful, thus a potential risk in medical environments. In contrast, reference-based LE approaches, directly manipulate the input image and analyse the resulting prediction's differences. Their basic idea is that an image-region's prediction-influence can be assessed by counterfactual reasoning: How would the prediction score vary, if the regions image-information would be missing, i.e. its contribution marginalized. The classically heuristic approaches, e.g. Gaussian noise and blurring or replacement by a predefined colour, have been advanced to local neighbourhood conditioning and stronger conditional generative models. Reference based LEs have the advantage of an a-priori clear and intuitively conveyable meaning of their result, hence address informativeness for end-users. However, in regard of medical images they come with a caveat. A prediction neutral region depicts per definition healthy tissue. Contradictory the presented approaches introduce noise and thereby possibly pathological indications or anatomically implausible tissue (cf. FIG. 1). By this, they raise the out-of-distribution problem: they expect a meaningful classification, based on an image outside the classifier's domain. Hence, they violate the needed faithfulness for clinical applications. Marginalization for medical imaging, that is generating sound, counterfactual images-regions, hence mapping between pathological tissue representations and their healthy equivalent, has become an active research area of its own. While generative adversarial networks (GANs) deliver promising results, their clinical-application is hindered by resolution limitation, and the same underlying out-of-distribution issue. In the state of the art the locally acting, globally conditioned, per-pixel reconstruction of partial convolution inpainting (PCI) is favoured over glo