Leveraging different learning styles for improved knowledge distillation in biomedical imaging.

Learning style refers to a type of training mechanism adopted by an individual to gain new knowledge. As suggested by the VARK model, humans have different learning preferences, like Visual (V), Auditory (A), Read/Write (R), and Kinesthetic (K), for acquiring and effectively processing information. Our work endeavors to leverage this concept of knowledge diversification to improve the performance of model compression techniques like Knowledge Distillation (KD) and Mutual Learning (ML). Consequently, we use a single-teacher and two-student network in a unified framework that not only allows for the transfer of knowledge from teacher to students (KD) but also encourages collaborative learning between students (ML). Unlike the conventional approach, where the teacher shares the same knowledge in the form of predictions or feature representations with the student network, our proposed approach employs a more diversified strategy by training one student with predictions and the other with feature maps from the teacher. We further extend this knowledge diversification by facilitating the exchange of predictions and feature maps between the two student networks, enriching their learning experiences. We have conducted comprehensive experiments with three benchmark datasets for both classification and segmentation tasks using two different network architecture combinations. These experimental results demonstrate that knowledge diversification in a combined KD and ML framework outperforms conventional KD or ML techniques (with similar network configuration) that only use predictions with an average improvement of 2%. Furthermore, consistent improvement in performance across different tasks, with various network architectures, and over state-of-the-art techniques establishes the robustness and generalizability of the proposed model.

Understanding calibration of deep neural networks for medical image classification.

Background and Objective - In the field of medical image analysis, achieving high accuracy is not enough; ensuring well-calibrated predictions is also crucial. Confidence scores of a deep neural network play a pivotal role in explainability by providing insights into the model's certainty, identifying cases that require attention, and establishing trust in its predictions. Consequently, the significance of a well-calibrated model becomes paramount in the medical imaging domain, where accurate and reliable predictions are of utmost importance. While there has been a significant effort towards training modern deep neural networks to achieve high accuracy on medical imaging tasks, model calibration and factors that affect it remain under-explored. Methods - To address this, we conducted a comprehensive empirical study that explores model performance and calibration under different training regimes. We considered fully supervised training, which is the prevailing approach in the community, as well as rotation-based self-supervised method with and without transfer learning, across various datasets and architecture sizes. Multiple calibration metrics were employed to gain a holistic understanding of model calibration. Results - Our study reveals that factors such as weight distributions and the similarity of learned representations correlate with the calibration trends observed in the models. Notably, models trained using rotation-based self-supervised pretrained regime exhibit significantly better calibration while achieving comparable or even superior performance compared to fully supervised models across different medical imaging datasets. Conclusion - These findings shed light on the importance of model calibration in medical image analysis and highlight the benefits of incorporating self-supervised learning approach to improve both performance and calibration.

Predicting women with depressive symptoms postpartum with machine learning methods.

Postpartum depression (PPD) is a detrimental health condition that affects 12% of new mothers. Despite negative effects on mothers' and children's health, many women do not receive adequate care. Preventive interventions are cost-efficient among high-risk women, but our ability to identify these is poor. We leveraged the power of clinical, demographic, and psychometric data to assess if machine learning methods can make accurate predictions of postpartum depression. Data were obtained from a population-based prospective cohort study in Uppsala, Sweden, collected between 2009 and 2018 (BASIC study, n = 4313). Sub-analyses among women without previous depression were performed. The extremely randomized trees method provided robust performance with highest accuracy and well-balanced sensitivity and specificity (accuracy 73%, sensitivity 72%, specificity 75%, positive predictive value 33%, negative predictive value 94%, area under the curve 81%). Among women without earlier mental health issues, the accuracy was 64%. The variables setting women at most risk for PPD were depression and anxiety during pregnancy, as well as variables related to resilience and personality. Future clinical models that could be implemented directly after delivery might consider including these variables in order to identify women at high risk for postpartum depression to facilitate individualized follow-up and cost-effectiveness.

Investigating the temporal dynamics of electroencephalogram (EEG) microstates using recurrent neural networks.

Electroencephalogram (EEG) microstates that represent quasi-stable, global neuronal activity are considered as the building blocks of brain dynamics. Therefore, the analysis of microstate sequences is a promising approach to understand fast brain dynamics that underlie various mental processes. Recent studies suggest that EEG microstate sequences are non-Markovian and nonstationary, highlighting the importance of the sequential flow of information between different brain states. These findings inspired us to model these sequences using Recurrent Neural Networks (RNNs) consisting of long-short-term-memory (LSTM) units to capture the complex temporal dependencies. Using an LSTM-based auto encoder framework and different encoding schemes, we modeled the microstate sequences at multiple time scales (200-2,000 ms) aiming to capture stably recurring microstate patterns within and across subjects. We show that RNNs can learn underlying microstate patterns with high accuracy and that the microstate trajectories are subject invariant at shorter time scales (≤400 ms) and reproducible across sessions. Significant drop in the reconstruction accuracy was observed for longer sequence lengths of 2,000 ms. These findings indirectly corroborate earlier studies which indicated that EEG microstate sequences exhibit long-range dependencies with finite memory content. Furthermore, we find that the latent representations learned by the RNNs are sensitive to external stimulation such as stress while the conventional univariate microstate measures (e.g., occurrence, mean duration, etc.) fail to capture such changes in brain dynamics. While RNNs cannot be configured to identify the specific discriminating patterns, they have the potential for learning the underlying temporal dynamics and are sensitive to sequence aberrations characterized by changes in metal processes. Empowered with the macroscopic understanding of the temporal dynamics that extends beyond short-term interactions, RNNs offer a reliable alternative for exploring system level brain dynamics using EEG microstate sequences.

REFUGE Challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs.

Glaucoma is one of the leading causes of irreversible but preventable blindness in working age populations. Color fundus photography (CFP) is the most cost-effective imaging modality to screen for retinal disorders. However, its application to glaucoma has been limited to the computation of a few related biomarkers such as the vertical cup-to-disc ratio. Deep learning approaches, although widely applied for medical image analysis, have not been extensively used for glaucoma assessment due to the limited size of the available data sets. Furthermore, the lack of a standardize benchmark strategy makes difficult to compare existing methods in a uniform way. In order to overcome these issues we set up the Retinal Fundus Glaucoma Challenge, REFUGE (https://refuge.grand-challenge.org), held in conjunction with MICCAI 2018. The challenge consisted of two primary tasks, namely optic disc/cup segmentation and glaucoma classification. As part of REFUGE, we have publicly released a data set of 1200 fundus images with ground truth segmentations and clinical glaucoma labels, currently the largest existing one. We have also built an evaluation framework to ease and ensure fairness in the comparison of different models, encouraging the development of novel techniques in the field. 12 teams qualified and participated in the online challenge. This paper summarizes their methods and analyzes their corresponding results. In particular, we observed that two of the top-ranked teams outperformed two human experts in the glaucoma classification task. Furthermore, the segmentation results were in general consistent with the ground truth annotations, with complementary outcomes that can be further exploited by ensembling the results.

Reward circuit connectivity relates to delay discounting in children with attention-deficit/hyperactivity disorder.

Attention-deficit/hyperactivity disorder (ADHD) is a prevalent psychiatric disorder that has poor long-term outcomes and remains a major public health concern. Recent theories have proposed that ADHD arises from alterations in multiple neural pathways. Alterations in reward circuits are hypothesized as one core dysfunction, leading to altered processing of anticipated rewards. The nucleus accumbens (NAcc) is particularly important for reward processes; task-based fMRI studies have found atypical activation of this region while the participants performed a reward task. Understanding how reward circuits are involved with ADHD may be further enhanced by considering how the NAcc interacts with other brain regions. Here we used the technique of resting-state functional connectivity MRI (rs-fcMRI) to examine the alterations in the NAcc interactions and how they relate to impulsive decision making in ADHD. Using rs-fcMRI, this study: examined differences in functional connectivity of the NAcc between children with ADHD and control children; correlated the functional connectivity of NAcc with impulsivity, as measured by a delay discounting task; and combined these two initial segments to identify the atypical NAcc connections that were associated with impulsive decision making in ADHD. We found that functional connectivity of NAcc was atypical in children with ADHD and the ADHD-related increased connectivity between NAcc and the prefrontal cortex was associated with greater impulsivity (steeper delayed-reward discounting). These findings are consistent with the hypothesis that atypical signaling of the NAcc to the prefrontal cortex in ADHD may lead to excessive approach and failure in estimating future consequences; thus, leading to impulsive behavior.

Bayesian analysis of fMRI data with ICA based spatial prior.

Spatial modeling is essential for fMRI analysis due to relatively high noise in the data. Earlier approaches have been primarily concerned with the spatial coherence of the BOLD response in local neighborhoods. In addition to a smoothness constraint, we propose to incorporate prior knowledge of brain activation patterns learned from training samples. This spatially informed prior can significantly enhance the estimation process by inducing sensitivity to task related regions of the brain. As fMRI data exhibits intersubject variability in functional anatomy, we design the prior using Independent Component Analysis (ICA). Due to the non-Gaussian assumption, ICA does not regress to the mean activation pattern and thus avoids suppressing intersubject differences. Results from a real fMRI experiment indicate that our approach provides statistically significant improvement in estimating activation compared to the standard general linear model (GLM) based methods.