Solving Inverse Problems With Deep Neural Networks - Robustness Included?

In the past five years, deep learning methods have become state-of-the-art in solving various inverse problems. Before such approaches can find application in safety-critical fields, a verification of their reliability appears mandatory. Recent works have pointed out instabilities of deep neural networks for several image reconstruction tasks. In analogy to adversarial attacks in classification, it was shown that slight distortions in the input domain may cause severe artifacts. The present article sheds new light on this concern, by conducting an extensive study of the robustness of deep-learning-based algorithms for solving underdetermined inverse problems. This covers compressed sensing with Gaussian measurements as well as image recovery from Fourier and Radon measurements, including a real-world scenario for magnetic resonance imaging (using the NYU-fastMRI dataset). Our main focus is on computing adversarial perturbations of the measurements that maximize the reconstruction error. A distinctive feature of our approach is the quantitative and qualitative comparison with total-variation minimization, which serves as a provably robust reference method. In contrast to previous findings, our results reveal that standard end-to-end network architectures are not only resilient against statistical noise, but also against adversarial perturbations. All considered networks are trained by common deep learning techniques, without sophisticated defense strategies.

Detecting failure modes in image reconstructions with interval neural network uncertainty.

PURPOSE: The quantitative detection of failure modes is important for making deep neural networks reliable and usable at scale. We consider three examples for common failure modes in image reconstruction and demonstrate the potential of uncertainty quantification as a fine-grained alarm system. METHODS: We propose a deterministic, modular and lightweight approach called Interval Neural Network (INN) that produces fast and easy to interpret uncertainty scores for deep neural networks. Importantly, INNs can be constructed post hoc for already trained prediction networks. We compare it against state-of-the-art baseline methods (MCDROP, PROBOUT). RESULTS: We demonstrate on controlled, synthetic inverse problems the capacity of INNs to capture uncertainty due to noise as well as directional error information. On a real-world inverse problem with human CT scans, we can show that INNs produce uncertainty scores which improve the detection of all considered failure modes compared to the baseline methods. CONCLUSION: Interval Neural Networks offer a promising tool to expose weaknesses of deep image reconstruction models and ultimately make them more reliable. The fact that they can be applied post hoc to equip already trained deep neural network models with uncertainty scores makes them particularly interesting for deployment.

How much information about the benefits of medicines is included in patient leaflets in the European Union? - A survey.

INTRODUCTION: Patient information leaflets (PILs) are required with all licensed medicines throughout the European Union (EU) and they must include information about all side effects and their likelihood. This has led to criticism of a lack of balance, with little information included about potential benefits. Recent European Medicines Agency guidance proposed the inclusion of benefit information, and this study examined the current prevalence and type of such information in PILs in the EU. METHODS: A survey and content analysis of the English translation of PILs in the EUwas carried out. Random quota sampling was used on the most frequently dispensed (n = 50) and newly licensed medicines (n = 50) in 2011/2. Leaflets were searched for benefit information meeting predefined criteria, and data synthesised and categorised into 10 categories. RESULTS: Eighty-five (85%) leaflets described how the medicine works, with 45 providing information about the rationale for treatment (more commonly for newly licensed (32/50) than most commonly dispensed medicines (13/50; P < 0.001). Nearly half (47) did not describe whether the medicine was curative, symptomatic or preventative. The terms used to communicate uncertainty were imprecise (such as 'may help'). None communicated numerical benefit information. CONCLUSION: Current PILs do not appropriately communicate information about benefit. At the basic level, around a half did not include information about treatment rationale or whether the treatment was to treat symptoms, curative or preventative. However, for true informed decision making, patients need quantitative information about benefits and none of the leaflets provided this.

Providing additional information about the benefits of statins in a leaflet for patients with coronary heart disease: a qualitative study of the impact on attitudes and beliefs.

OBJECTIVE: To explore the impact of providing additional information about the potential benefits of simvastatin in a patient leaflet on attitudes and beliefs. DESIGN: Interview-based study using a generic qualitative approach and framework analysis. PARTICIPANTS: 21 participants receiving a prescription for simvastatin were recruited from a general practitioner practice (from a total of 120). 8 participants were women; the age range was 55-92. INTERVENTION: Participants were provided with leaflets showing one of 3 types of additional benefit information: (1) textual statement, (2) number needed to treat (NNT) or (3) natural frequency. Semistructured interviews explored patient's attitudes and beliefs. RESULTS: A descriptive narrative of preferences for format suggested patients prefer textual as opposed to numerical benefit information. Significant barriers to the acceptance of numerical benefit information included difficulty in understanding the numbers. Patients overestimated the benefits of statins and expressed surprise at the numerical information. CONCLUSIONS: Textual information was preferred but numerical information, in particular in the form of a natural frequency, may help patients make judgements about their medicines. NNTs were found to be very difficult to understand. This raises the prospect that some patients might reject medicines because of disappointment with the perceived low benefits of their medicines. The self-reported impact on behaviour appeared minimal with reports of intentions to 'do what the doctor tells me'. Further research is needed to explore the impact of such statements on people who are yet to be prescribed a statin.

Do Patients Use a Headline Section in a Leaflet to Find Key Information About Their Medicines? Findings From a User-Test Study.

BACKGROUND: In the European Union (EU), all medicines are mandated to be provided with a patient information leaflet (PIL). Many patients express concerns about the length and complexity of some PILs, and this can be a disincentive for patients to read the PILS. In order to address this, the UK's regulatory body (Medicine and Healthcare products Regulatory Agency [MHRA]) suggested leaflets might include a headline section-information presented prominently at the beginning of a leaflet that summarizes key safety messages about a drug. OBJECTIVE: To explore the extent to which readers used a headline section in a PIL, using a form of diagnostic testing called user-testing, which examines how readers find and understand key information. METHODS: The study used a cross-sectional design to user-test a PIL with a headline section in a target sample of 20 participants. Participants were provided with an exemplar PIL, and the performance of the PIL was evaluated by a questionnaire and semistructured interview. RESULTS: The results showed that a headline section was used just over one-third of the time (39%); 90% of participants used the headline section to find information when they initially began the user-test. The qualitative findings suggested that the participants valued the presence of the headline section. CONCLUSION: The research suggests there does not appear to be any negative impact from including a headline section in a PIL, and it is a technique that is highly valued by the consumers of medicines information.

Changes in Side Effect Risk Communication in Patient Information Leaflets over the Past Decade: Results of a Survey.

INTRODUCTION: Patients' perceptions of side effect risks are important influences on their medicine-taking behaviour. A previous survey of Patient Information Leaflets (PILs) showed considerable variation in the terms used to communicate risks. OBJECTIVE: Our objective was to assess the methods used to describe risk of side effects in recent PILs and to compare them with PILs sampled in 2006. METHOD: We sampled PILs for the 50 most frequently dispensed medicines in England and Wales in 2012 and PILs for the 50 most recently licensed medicines. We assessed the use of risk frequency terms or numbers, and the use of the risk format recommended by the European Medicines Agency (EMA). RESULTS: A majority (76 %) of PILs for the most frequently dispensed medicines included a risk frequency descriptor, with 66 % using the recommended format. No difference was seen between PILs for branded and generic medicines. All 50 PILs for the most recently licensed medicines used the EU recommended risk format. PILs from the 2012 sample were much more likely than those from the 2006 sample to include risk descriptors and to use a consistent approach. CONCLUSION: The increased use and consistency of risk descriptors in PILs should benefit patients, particularly those using multiple medicines produced by different market authorisation holders. A need remains for further research evaluating the risk format recommended by the EMA. There is also a need for research evaluating spoken information and other sources of printed risk information about medicines that is available to patients.

Noninvasive imaging of cell death using an Hsp90 ligand.

Cell death plays a central role in normal physiology and in disease. Common to apoptotic and necrotic cell death is the eventual loss of plasma membrane integrity. We have produced a small organoarsenical compound, 4-(N-(S-glutathionylacetyl)amino)phenylarsonous acid, that rapidly accumulates in the cytosol of dying cells coincident with loss of plasma membrane integrity. The compound is retained in the cytosol predominantly by covalent reaction with the 90 kDa heat shock protein (Hsp90), the most abundant molecular chaperone of the eukaryotic cytoplasm. The organoarsenical was tagged with either optical or radioisotope reporting groups to image cell death in cultured cells and in murine tumors ex vivo and in situ. Tumor cell death in mice was noninvasively imaged by SPECT/CT using an (111)In-tagged compound. This versatile compound should enable the imaging of cell death in most experimental settings.