Detection of fibrosing interstitial lung disease-suspected chest radiographs using a deep learning-based computer-aided detection system: a retrospective, observational study.

OBJECTIVES: To investigate the effectiveness of BMAX, a deep learning-based computer-aided detection system for detecting fibrosing interstitial lung disease (ILD) on chest radiographs among non-expert and expert physicians in the real-world clinical setting. DESIGN: Retrospective, observational study. SETTING: This study used chest radiograph images consecutively taken in three medical facilities with various degrees of referral. Three expert ILD physicians interpreted each image and determined whether it was a fibrosing ILD-suspected image (fibrosing ILD positive) or not (fibrosing ILD negative). Interpreters, including non-experts and experts, classified each of 120 images extracted from the pooled data for the reading test into positive or negative for fibrosing ILD without and with the assistance of BMAX. PARTICIPANTS: Chest radiographs of patients aged 20 years or older with two or more visits that were taken during consecutive periods were accumulated. 1251 chest radiograph images were collected, from which 120 images (24 positive and 96 negative images) were randomly extracted for the reading test. The interpreters for the reading test were 20 non-expert physicians and 5 expert physicians (3 pulmonologists and 2 radiologists). PRIMARY AND SECONDARY OUTCOME MEASURES: The primary outcome was the comparison of area under the receiver-operating characteristic curve (ROC-AUC) for identifying fibrosing ILD-positive images by non-experts without versus with BMAX. The secondary outcome was the comparison of sensitivity, specificity and accuracy by non-experts and experts without versus with BMAX. RESULTS: The mean ROC-AUC of non-expert interpreters was 0.795 (95% CI; 0.765 to 0.825) without BMAX and 0.825 (95% CI; 0.799 to 0.850) with BMAX (p=0.005). After using BMAX, sensitivity was improved from 0.744 (95% CI; 0.697 to 0.791) to 0.802 (95% CI; 0.754 to 0.850) among non-experts (p=0.003), but not among experts (p=0.285). Specificity and accuracy were not changed after using BMAX among either non-expert or expert interpreters. CONCLUSION: BMAX was useful for detecting fibrosing ILD-suspected chest radiographs for non-expert physicians. TRIAL REGISTRATION NUMBER: jRCT1032220090.

Adversarial attacks and defenses using feature-space stochasticity.

Recent studies in deep neural networks have shown that injecting random noise in the input layer of the networks contributes towards ℓp-norm-bounded adversarial perturbations. However, to defend against unrestricted adversarial examples, most of which are not ℓp-norm-bounded in the input layer, such input-layer random noise may not be sufficient. In the first part of this study, we generated a novel class of unrestricted adversarial examples termed feature-space adversarial examples. These examples are far from the original data in the input space but adjacent to the original data in a hidden-layer feature space and far again in the output layer. In the second part of this study, we empirically showed that while injecting random noise in the input layer was unable to defend these feature-space adversarial examples, they were defended by injecting random noise in the hidden layer. These results highlight the novel benefit of stochasticity in higher layers, in that it is useful for defending against these feature-space adversarial examples, a class of unrestricted adversarial examples.

Causal importance of low-level feature selectivity for generalization in image recognition.

Although our brain and deep neural networks (DNNs) can perform high-level sensory-perception tasks, such as image or speech recognition, the inner mechanism of these hierarchical information-processing systems is poorly understood in both neuroscience and machine learning. Recently, Morcos et al. (2018) examined the effect of class-selective units in DNNs, i.e., units with high-level selectivity, on network generalization, concluding that hidden units that are selectively activated by specific input patterns may harm the network's performance. In this study, we revisited their hypothesis, considering units with selectivity for lower-level features, and argue that selective units are not always harmful to the network performance. Specifically, by using DNNs trained for image classification, we analyzed the orientation selectivity of individual units, a low-level selectivity widely studied in visual neuroscience. We found that orientation-selective units exist in both lower and higher layers of these DNNs, as in our brain. In particular, units in lower layers became more orientation-selective as the generalization performance improved during the course of training. Consistently, networks that generalized better were more orientation-selective in the lower layers. We finally revealed that ablating these selective units in the lower layers substantially degraded the generalization performance of the networks, at least by disrupting the shift-invariance of the higher layers. These results suggest that orientation selectivity can play a causally important role in object recognition, and that, contrary to the triviality of units with high-level selectivity, lower-layer units with selectivity for low-level features may be indispensable for generalization, at least for the several network architectures.

Characterisation of nonlinear receptive fields of visual neurons by convolutional neural network.

A comprehensive understanding of the stimulus-response properties of individual neurons is necessary to crack the neural code of sensory cortices. However, a barrier to achieving this goal is the difficulty of analysing the nonlinearity of neuronal responses. Here, by incorporating convolutional neural network (CNN) for encoding models of neurons in the visual cortex, we developed a new method of nonlinear response characterisation, especially nonlinear estimation of receptive fields (RFs), without assumptions regarding the type of nonlinearity. Briefly, after training CNN to predict the visual responses to natural images, we synthesised the RF image such that the image would predictively evoke a maximum response. We first demonstrated the proof-of-principle using a dataset of simulated cells with various types of nonlinearity. We could visualise RFs with various types of nonlinearity, such as shift-invariant RFs or rotation-invariant RFs, suggesting that the method may be applicable to neurons with complex nonlinearities in higher visual areas. Next, we applied the method to a dataset of neurons in mouse V1. We could visualise simple-cell-like or complex-cell-like (shift-invariant) RFs and quantify the degree of shift-invariance. These results suggest that CNN encoding model is useful in nonlinear response analyses of visual neurons and potentially of any sensory neurons.

Efficient Prediction of Vitamin B Deficiencies via Machine-Learning Using Routine Blood Test Results in Patients With Intense Psychiatric Episode.

BACKGROUND: Vitamin B deficiency is common worldwide and may lead to psychiatric symptoms; however, vitamin B deficiency epidemiology in patients with intense psychiatric episode has rarely been examined. Moreover, vitamin deficiency testing is costly and time-consuming, which has hampered effectively ruling out vitamin deficiency-induced intense psychiatric symptoms. In this study, we aimed to clarify the epidemiology of these deficiencies and efficiently predict them using machine-learning models from patient characteristics and routine blood test results that can be obtained within one hour. METHODS: We reviewed 497 consecutive patients, who are deemed to be at imminent risk of seriously harming themselves or others, over a period of 2 years in a single psychiatric tertiary-care center. Machine-learning models (k-nearest neighbors, logistic regression, support vector machine, and random forest) were trained to predict each deficiency from age, sex, and 29 routine blood test results gathered in the period from September 2015 to December 2016. The models were validated using a dataset collected from January 2017 through August 2017. RESULTS: We found that 112 (22.5%), 80 (16.1%), and 72 (14.5%) patients had vitamin B1, vitamin B12, and folate (vitamin B9) deficiency, respectively. Further, the machine-learning models were well generalized to predict deficiency in the future unseen data, especially using random forest; areas under the receiver operating characteristic curves for the validation dataset (i.e., the dataset not used for training the models) were 0.716, 0.599, and 0.796, respectively. The Gini importance of these vitamins provided further evidence of a relationship between these vitamins and the complete blood count, while also indicating a hitherto rarely considered, potential association between these vitamins and alkaline phosphatase (ALP) or thyroid stimulating hormone (TSH). DISCUSSION: This study demonstrates that machine-learning can efficiently predict some vitamin deficiencies in patients with active psychiatric symptoms, based on the largest cohort to date with intense psychiatric episode. The prediction method may expedite risk stratification and clinical decision-making regarding whether replacement therapy should be prescribed. Further research includes validating its external generalizability in other clinical situations and clarify whether interventions based on this method could improve patient care and cost-effectiveness.