Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

Ophthalmology Ophthalmology

Differentiating malignant and benign eyelid lesions using deep learning.

In Scientific reports ; h5-index 158.0

Artificial intelligence as a screening tool for eyelid lesions will be helpful for early diagnosis of eyelid malignancies and proper decision-making. This study aimed to evaluate the performance of a deep learning model in differentiating eyelid lesions using clinical eyelid photographs in comparison with human ophthalmologists. We included 4954 photographs from 928 patients in this retrospective cross-sectional study. Images were classified into three categories: malignant lesion, benign lesion, and no lesion. Two pre-trained convolutional neural network (CNN) models, DenseNet-161 and EfficientNetV2-M architectures, were fine-tuned to classify images into three or two (malignant versus benign) categories. For a ternary classification, the mean diagnostic accuracies of the CNNs were 82.1% and 83.0% using DenseNet-161 and EfficientNetV2-M, respectively, which were inferior to those of the nine clinicians (87.0-89.5%). For the binary classification, the mean accuracies were 87.5% and 92.5% using DenseNet-161 and EfficientNetV2-M models, which was similar to that of the clinicians (85.8-90.0%). The mean AUC of the two CNN models was 0.908 and 0.950, respectively. Gradient-weighted class activation map successfully highlighted the eyelid tumors on clinical photographs. Deep learning models showed a promising performance in discriminating malignant versus benign eyelid lesions on clinical photographs, reaching the level of human observers.

Lee Min Joung, Yang Min Kyu, Khwarg Sang In, Oh Eun Kyu, Choi Youn Joo, Kim Namju, Choung Hokyung, Seo Chang Won, Ha Yun Jong, Cho Min Ho, Cho Bum-Joo

2023-Mar-13

Surgery Surgery

An artificial intelligence approach for predicting death or organ failure after hospitalization for COVID-19: development of a novel risk prediction tool and comparisons with ISARIC-4C, CURB-65, qSOFA, and MEWS scoring systems.

In Respiratory research ; h5-index 45.0

BACKGROUND : We applied machine learning (ML) algorithms to generate a risk prediction tool [Collaboration for Risk Evaluation in COVID-19 (CORE-COVID-19)] for predicting the composite of 30-day endotracheal intubation, intravenous administration of vasopressors, or death after COVID-19 hospitalization and compared it with the existing risk scores.

METHODS : This is a retrospective study of adults hospitalized with COVID-19 from March 2020 to February 2021. Patients, each with 92 variables, and one composite outcome underwent feature selection process to identify the most predictive variables. Selected variables were modeled to build four ML algorithms (artificial neural network, support vector machine, gradient boosting machine, and Logistic regression) and an ensemble model to generate a CORE-COVID-19 model to predict the composite outcome and compared with existing risk prediction scores. The net benefit for clinical use of each model was assessed by decision curve analysis.

RESULTS : Of 1796 patients, 278 (15%) patients reached primary outcome. Six most predictive features were identified. Four ML algorithms achieved comparable discrimination (P > 0.827) with c-statistics ranged 0.849-0.856, calibration slopes 0.911-1.173, and Hosmer-Lemeshow P > 0.141 in validation dataset. These 6-variable fitted CORE-COVID-19 model revealed a c-statistic of 0.880, which was significantly (P < 0.04) higher than ISARIC-4C (0.751), CURB-65 (0.735), qSOFA (0.676), and MEWS (0.674) for outcome prediction. The net benefit of the CORE-COVID-19 model was greater than that of the existing risk scores.

CONCLUSION : The CORE-COVID-19 model accurately assigned 88% of patients who potentially progressed to 30-day composite events and revealed improved performance over existing risk scores, indicating its potential utility in clinical practice.

Kwok Stephen Wai Hang, Wang Guanjin, Sohel Ferdous, Kashani Kianoush B, Zhu Ye, Wang Zhen, Antpack Eduardo, Khandelwal Kanika, Pagali Sandeep R, Nanda Sanjeev, Abdalrhim Ahmed D, Sharma Umesh M, Bhagra Sumit, Dugani Sagar, Takahashi Paul Y, Murad Mohammad H, Yousufuddin Mohammed

2023-Mar-13

COVID-19, Machine learning algorithms, Mortality, Organ failure, Prediction models

General General

Multi-Modal Facial Expression Recognition with Transformer-Based Fusion Networks and Dynamic Sampling

ArXiv Preprint

Facial expression recognition is important for various purpose such as emotion detection, mental health analysis, and human-machine interaction. In facial expression recognition, incorporating audio information along with still images can provide a more comprehensive understanding of an expression state. This paper presents the Multi-modal facial expression recognition methods for Affective Behavior in-the-wild (ABAW) challenge at CVPR 2023. We propose a Modal Fusion Module (MFM) to fuse audio-visual information. The modalities used are image and audio, and features are extracted based on Swin Transformer to forward the MFM. Our approach also addresses imbalances in the dataset through data resampling in training dataset and leverages the rich modal in a single frame using dynmaic data sampling, leading to improved performance.

Jun-Hwa Kim, Namho Kim, Chee Sun Won

2023-03-15

Radiology Radiology

Development and validation of a prognostic model for HER2-low breast cancer to evaluate neoadjuvant therapy.

In Gland surgery

BACKGROUND : Human epidermal growth factor receptor 2 (HER2) low breast cancer (BC) accounts for 30-51% of all BCs. How to precisely assess the response to neoadjuvant therapy in this heterogenous tumor is currently unanswered. With the advance in multi-omics, refining the molecular subtyping other than the current hormone receptor (HR)-based subtyping to guide the neoadjuvant therapy for HER2-low BC is potentially feasible.

METHODS : The messenger RNA (mRNA), clinical, and pathological data of all HER2-low BC patients (n=368) from the Neoadjuvant I-SPY2 Trial, were retrieved. Ninety-eight patients achieved pathological complete response (pCR) were randomly divided into the training and validation sets with 8:2 ratio. The non-pCR cases were corporated into the above datasets with 1:1 ratio. The rest non-pCR cases were served as the test set. Random forest (RF), support vector machine (SVM), and fully connected neural network (FCNN) were applied to establish a 1-dimensional (1D) model based on mRNA data. The method with best prediction value among the 3 models was selected for further modeling when combining pathological features. A new classification of deep learning (CDn) was proposed based on a multi-omics model. After identifying pCR-related features by the integral gradient and unsupervised hierarchical clustering method, the responses to neoadjuvant therapy associated with these features across different subgroups were analyzed.

RESULTS : Compared with the RF and SVM models, the FCNN model achieved the best performance [area under the curve (AUC): 0.89] based on the mRNA feature. By combining mRNA and pathological features, the FCNN model proposed 2 new subtypes including CD1 and CD0 for HER2-low BC. CD1 increased the sensitivity to predict pCR by 23.5% [to 87.8%; 95% confidence interval (CI): 78% to 94%] and improved the specificity to pCR by 12.2% (to 77.4%; 95% CI: 69% to 87%) when comparing with the current HR classification for HER2-low BC.

CONCLUSIONS : The new typing method (CD1 and CD0) proposed in this study achieved excellent performance for predicting the pCR to neoadjuvant therapy in HER2-low BC. The patients who were not sensitive to neoadjuvant therapy according to multi-omics models might receive surgical treatment directly.

Li Xiaoping, Lin Zhiquan, Yu Qihe, Qiu Chaoran, Lai Chan, Huang Hui, Zhang Yiwen, Zhang Weibin, Zhu Jintao, Huang Xin, Li Weiwen

2023-Feb-28

Breast cancer (BC), deep learning, human epidermal growth factor receptor 2-low (HER2-low), integral gradient, molecular subtype

Surgery Surgery

Marmoset monkeys use different avoidance strategies to cope with ambient noise during vocal behavior.

In iScience

Multiple strategies have evolved to compensate for masking noise, leading to changes in call features. One call adjustment is the Lombard effect, an increase in call amplitude in response to noise. Another strategy involves call production in periods where noise is absent. While mechanisms underlying vocal adjustments have been well studied, mechanisms underlying noise avoidance strategies remain largely unclear. We systematically perturbed ongoing phee calls of marmosets to investigate noise avoidance strategies. Marmosets canceled their calls after noise onset and produced longer calls after noise-phases ended. Additionally, the number of uttered syllables decreased during noise perturbation. This behavior persisted beyond the noise-phase. Using machine learning techniques, we found that a fraction of single phees were initially planned as double phees and became interrupted after the first syllable. Our findings indicate that marmosets use different noise avoidance strategies and suggest vocal flexibility at different complexity levels in the marmoset brain.

Löschner Julia, Pomberger Thomas, Hage Steffen R

2023-Mar-17

Behavioral neuroscience, Biological sciences, Machine learning

General General

Predicting motor behavior: an efficient EEG signal processing pipeline to detect brain states with potential therapeutic relevance for VR-based neurorehabilitation.

In Virtual reality

Virtual reality (VR)-based motor therapy is an emerging approach in neurorehabilitation. The combination of VR with electroencephalography (EEG) presents further opportunities to improve therapeutic efficacy by personalizing the paradigm. Specifically, the idea is to synchronize the choice and timing of stimuli in the perceived virtual world with fluctuating brain states relevant to motor behavior. Here, we present an open source EEG single-trial based classification pipeline that is designed to identify ongoing brain states predictive of the planning and execution of movements. 9 healthy volunteers each performed 1080 trials of a repetitive reaching task with an implicit two-alternative forced choice, i.e., use of the right or left hand, in response to the appearance of a visual target. The performance of the EEG decoding pipeline was assessed with respect to classification accuracy of right vs. left arm use, based on the EEG signal at the time of the stimulus. Different features, feature extraction methods, and classifiers were compared at different time windows; the number and location of informative EEG channels and the number of calibration trials needed were also quantified, as well as any benefits from individual-level optimization of pipeline parameters. This resulted in a set of recommended parameters that achieved an average 83.3% correct prediction on never-before-seen testing data, and a state-of-the-art 77.1% in a real-time simulation. Neurophysiological plausibility of the resulting classifiers was assessed by time-frequency and event-related potential analyses, as well as by Independent Component Analysis topographies and cortical source localization. We expect that this pipeline will facilitate the identification of relevant brain states as prospective therapeutic targets in closed-loop EEG-VR motor neurorehabilitation.

McDermott Eric J, Metsomaa Johanna, Belardinelli Paolo, Grosse-Wentrup Moritz, Ziemann Ulf, Zrenner Christoph

2023

Brain-state decoding, Brain–computer interface (BCI), Classification, EEG, EEG-VR, Hand selection, Human-in-the-loop, Machine learning, Motor behavior, Motor intention, Neurorehabilitation, Open source pipeline, Pre-movement, Virtual reality