Dermatologist-like explainable AI enhances trust and confidence in diagnosing melanoma.

Artificial intelligence (AI) systems have been shown to help dermatologists diagnose melanoma more accurately, however they lack transparency, hindering user acceptance. Explainable AI (XAI) methods can help to increase transparency, yet often lack precise, domain-specific explanations. Moreover, the impact of XAI methods on dermatologists' decisions has not yet been evaluated. Building upon previous research, we introduce an XAI system that provides precise and domain-specific explanations alongside its differential diagnoses of melanomas and nevi. Through a three-phase study, we assess its impact on dermatologists' diagnostic accuracy, diagnostic confidence, and trust in the XAI-support. Our results show strong alignment between XAI and dermatologist explanations. We also show that dermatologists' confidence in their diagnoses, and their trust in the support system significantly increase with XAI compared to conventional AI. This study highlights dermatologists' willingness to adopt such XAI systems, promoting future use in the clinic.

Uncertainty Estimation in Medical Image Classification: Systematic Review.

BACKGROUND: Deep neural networks are showing impressive results in different medical image classification tasks. However, for real-world applications, there is a need to estimate the network's uncertainty together with its prediction. OBJECTIVE: In this review, we investigate in what form uncertainty estimation has been applied to the task of medical image classification. We also investigate which metrics are used to describe the effectiveness of the applied uncertainty estimation. METHODS: Google Scholar, PubMed, IEEE Xplore, and ScienceDirect were screened for peer-reviewed studies, published between 2016 and 2021, that deal with uncertainty estimation in medical image classification. The search terms "uncertainty," "uncertainty estimation," "network calibration," and "out-of-distribution detection" were used in combination with the terms "medical images," "medical image analysis," and "medical image classification." RESULTS: A total of 22 papers were chosen for detailed analysis through the systematic review process. This paper provides a table for a systematic comparison of the included works with respect to the applied method for estimating the uncertainty. CONCLUSIONS: The applied methods for estimating uncertainties are diverse, but the sampling-based methods Monte-Carlo Dropout and Deep Ensembles are used most frequently. We concluded that future works can investigate the benefits of uncertainty estimation in collaborative settings of artificial intelligence systems and human experts. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): RR2-10.2196/11936.

Explainable artificial intelligence in skin cancer recognition: A systematic review.

BACKGROUND: Due to their ability to solve complex problems, deep neural networks (DNNs) are becoming increasingly popular in medical applications. However, decision-making by such algorithms is essentially a black-box process that renders it difficult for physicians to judge whether the decisions are reliable. The use of explainable artificial intelligence (XAI) is often suggested as a solution to this problem. We investigate how XAI is used for skin cancer detection: how is it used during the development of new DNNs? What kinds of visualisations are commonly used? Are there systematic evaluations of XAI with dermatologists or dermatopathologists? METHODS: Google Scholar, PubMed, IEEE Explore, Science Direct and Scopus were searched for peer-reviewed studies published between January 2017 and October 2021 applying XAI to dermatological images: the search terms histopathological image, whole-slide image, clinical image, dermoscopic image, skin, dermatology, explainable, interpretable and XAI were used in various combinations. Only studies concerned with skin cancer were included. RESULTS: 37 publications fulfilled our inclusion criteria. Most studies (19/37) simply applied existing XAI methods to their classifier to interpret its decision-making. Some studies (4/37) proposed new XAI methods or improved upon existing techniques. 14/37 studies addressed specific questions such as bias detection and impact of XAI on man-machine-interactions. However, only three of them evaluated the performance and confidence of humans using CAD systems with XAI. CONCLUSION: XAI is commonly applied during the development of DNNs for skin cancer detection. However, a systematic and rigorous evaluation of its usefulness in this scenario is lacking.

Effects of Aging and Dual-Task Demands on the Comprehension of Less Expected Sentence Continuations: Evidence From Pupillometry.

Prior studies on language processing in aging have shown that older adults experience integration difficulties for contextually unpredictable target words (as indicated by low cloze probabilities in prior ratings), and that such comprehension difficulties are more likely to occur under more demanding processing conditions (e.g., dual-task situations). However, these effects have primarily been demonstrated for conditions when cloze probability of the linguistic stimuli was very low. The question we asked here was do dual-task demands also impair comprehension when target words provide a good, but not perfect, match with prior context? We used a dual-task design, consisting of a sentence comprehension and secondary motor tracking task. Critical target words were those which were not perfectly predictable based on context (words with a cloze probability of 0.7), as opposed to words that were near perfectly predictable based on context (cloze probabilities of 0.99). As a measure to index online processing difficulty for less expected target words, we took into account participants' pupil size. Separate mixed effects models were fit for language comprehension, motor tracking, and pupil size, showing the following: (1) dual-task demands led to age-related comprehension difficulties when target words were less expected (as opposed to very highly expected), (2) integration difficulty in older adults was related to cognitive overload as less expected sentence continuations progressed over time, resulting in behavioral trade-offs between language comprehension and motor tracking, and (3) lower levels of working memory were predictive of whether or not older adults experienced cognitive overload when processing less expected words. In sum, more demanding processing conditions lead to comprehension impairments when words are highly unpredictable based on context, as many prior studies showed. Comprehension impairments among older individuals also occur for conditions when words provide a good, but not perfect, match with prior context. Higher working memory capacity can alleviate such impairments in older adults, thereby suggesting that only high-WM older adults have sufficient cognitive resources to pre-activate words that complete a sentence context plausibly, but not perfectly.

Surprisal modulates dual-task performance in older adults: Pupillometry shows age-related trade-offs in task performance and time-course of language processing.

Even though older adults are known to have difficulty at language processing when a secondary task has to be performed simultaneously, few studies have addressed how older adults process language in dual-task demands when linguistic load is systematically varied. Here, we manipulated surprisal, an information theoretic measure that quantifies the amount of new information conveyed by a word, to investigate how linguistic load affects younger and older adults during early and late stages of sentence processing under conditions when attention is split between two tasks. In high-surprisal sentences, target words were implausible and mismatched with semantic expectancies based on context, thereby causing integration difficulty. Participants performed semantic meaningfulness judgments on sentences that were presented in isolation (single task) or while performing a secondary tracking task (dual task). Cognitive load was measured by means of pupillometry. Mixed-effects models were fit to the data, showing the following: (a) During the dual task, younger but not older adults demonstrated early sensitivity to surprisal (higher levels of cognitive load, indexed by pupil size) as sentences were heard online; (b) Older adults showed no immediate reaction to surprisal, but a delayed response, where their meaningfulness judgments to high-surprisal words remained stable in accuracy, while secondary tracking performance declined. Findings are discussed in relation to age-related trade-offs in dual tasking and differences in the allocation of attentional resources during language processing. Collectively, our data show that higher linguistic load leads to task trade-offs in older adults and differently affects the time course of online language processing in aging. (PsycINFO Database Record (c) 2018 APA, all rights reserved).

The role of the ventro-lateral prefrontal cortex in idiom comprehension: An rTMS study.

Previous research is equivocal with respect to the neural substrates of idiom processing. Particularly elusive is the role of the left ventro-lateral prefrontal cortex (VLPFC), a region implicated in semantic control generally. Although fMRI studies have shown that the left VLPFC is active during idiom processing (see Rapp et al. (2012), for review), rTMS studies have failed to corroborate a clear role of this prefrontal region (e.g., Oliveri et al., 2004). We investigated this issue using a semantic meaningfulness judgment task that compared idiom comprehension following rTMS-stimulation to the left VLPFC relative to a control site (vertex). We also investigated whether individual differences in general cognitive capacity among comprehenders modulated the effects of rTMS. The results indicate that left VLPFC stimulation particularly affected the processing of low-familiar idioms, possibly because these items involve a maximal semantic conflict between a salient literal and less-known figurative meaning. Of note, this pattern only emerged for comprehenders with higher cognitive control capacity, possibly because they were more likely to activate or maintain multiple semantic representations during idiom processing, which required VLPFC integrity. Taken together, the results support the importance of the left VLPFC to idiom processing.

Functional lateralization of lexical stress representation: a systematic review of patient data.

According to the functional lateralization hypothesis (FLH) the lateralization of speech prosody depends both on its function (linguistic = left, emotional = right) and on the size of the units it operates on (small = left, large = right). In consequence, according to the FLH, lexical stress should be processed by the left (language-dominant) hemisphere, given its linguistic function and small unit size. We performed an exhaustive search for case studies of patients with acquired dysprosody due to unilateral brain damage. In contrast to previous reviews we only regarded dysprosody at the lexical level (excluding phrasal stress). Moreover, we focused on the representational stage of lexical stress processing, excluding more peripheral perceptual or motor deficits. Applying these criteria, we included nine studies reporting on 11 patients. All of these patients showed representational deficits in word stress processing following a lesion in their language-dominant hemisphere. In 9 out of 11 patients, it was the left hemisphere which was affected. This is a much more consistent pattern as found in previous reviews, in which less rigorous inclusion criteria may have blurred the pattern of results. We conclude that the representation of lexical stress crucially relies on the functioning of the language-dominant (mostly left) hemisphere.

Individual differences in executive control relate to metaphor processing: an eye movement study of sentence reading.

Metaphors are common elements of language that allow us to creatively stretch the limits of word meaning. However, metaphors vary in their degree of novelty, which determines whether people must create new meanings on-line or retrieve previously known metaphorical meanings from memory. Such variations affect the degree to which general cognitive capacities such as executive control are required for successful comprehension. We investigated whether individual differences in executive control relate to metaphor processing using eye movement measures of reading. Thirty-nine participants read sentences including metaphors or idioms, another form of figurative language that is more likely to rely on meaning retrieval. They also completed the AX-CPT, a domain-general executive control task. In Experiment 1, we examined sentences containing metaphorical or literal uses of verbs, presented with or without prior context. In Experiment 2, we examined sentences containing idioms or literal phrases for the same participants to determine whether the link to executive control was qualitatively similar or different to Experiment 1. When metaphors were low familiar, all people read verbs used as metaphors more slowly than verbs used literally (this difference was smaller for high familiar metaphors). Executive control capacity modulated this pattern in that high executive control readers spent more time reading verbs when a prior context forced a particular interpretation (metaphorical or literal), and they had faster total metaphor reading times when there was a prior context. Interestingly, executive control did not relate to idiom processing for the same readers. Here, all readers had faster total reading times for high familiar idioms than literal phrases. Thus, executive control relates to metaphor but not idiom processing for these readers, and for the particular metaphor and idiom reading manipulations presented.