The ability to analyze data is enhanced, when it is represented in a conducive manner. Here, we report an autoencoder-based content-aware 2D representation of 1D time-series. We propose a novel representation of time-series utilizing both time- and frequency-domain analyses, carried out separately but concurrently. The loss function designed for autoencoder training ensures that the learnt latent space representation comprises of time-invariant features. Every element of the time-series is represented as a tuple with two components, one each, from latent space representation in time- and frequency-domains. An enhanced time-frequency domain representation is constructed whose axes are the latent spaces. In this representation, those tuples that represent the points in the time-series, together form the “Latent Space Signature” (LSS) of the input time-series. The obtained binary LSS’s can be processed for classification, saliency determination, etc. We illustrate the efficacy of LSS technique for classification, on 2 scenarios (with publicly available, bench-marked datasets) which are ECG time-series (109446 (5 labels) + 14552 (2 labels)) and Black hole time-series (astronomy data, consisting of 12 temporal classes, associated with 2 labels). Obtained results are compared against those of popular techniques, with concurring or improved results, that illustrate promise in the proposed technique.

We propose a novel approach to quantify "deviation from stochasticity" (DS) in a time-series. This is important to determine if the time-series is coming from a physical phenomenon or if it is noise. This approach utilizes time-invariant representation obtained using time- and frequency-domain analyses. Autoencoder based time-invariant features have been utilized to obtain multi-scale reconstruction as well as identification of prominent peaks in dissimilarity curves. We devise a DS measure based on the observation that a stochastic time-series exhibits similar behavior across multiple time scales. The values of DS are expected to be significantly small for stochastic time-series in comparison with those for non-stochastic time-series, leading to classification. As proof of concept, we illustrate this trend on synthetic data. Subsequently, the proposed methodology is applied on astronomical data which are 12 distinct temporal classes of time-series pertaining to the black hole GRS 1915 + 105, obtained from RXTE satellite. This dataset had been previously studied using correlation integration (CI) based approach to understand the underlying dynamics leading to time-series classification. Results obtained using the proposed methodology are compared with those obtained using CI. Concurrence is obtained for 11 temporal classes, while one is found to be non-concurrent. This could be attributed to the observation that the non-concurrence is due to that specific time-series exhibiting both stochastic and non-stochastic characteristics. Besides, these DS values can also be interpreted as quantification of signal-to-noise ratio (SNR) of a time-series.

Functional MRI (fMRI) is widely used to examine brain functionality by detecting alteration in oxygenated blood flow that arises with brain activity. In this study, complexity-specific image categorization across different visual datasets is performed using fMRI time series (TS) to understand differences in neuronal activities related to vision. Publicly available BOLD5000 dataset is used for this purpose, containing fMRI scans while viewing 5254 images of diverse categories, drawn from three standard computer vision datasets: COCO, ImageNet and SUN. To understand vision, it is important to study how brain functions while looking at different images. To achieve this, spatial encoding of fMRI BOLD TS has been performed that uses classical Gramian Angular Field (GAF) and Markov Transition Field (MTF) to obtain 2D BOLD TS, representing images of COCO, Imagenet and SUN. For classification, individual GAF and MTF features are fed into regular CNN. Subsequently, parallel CNN model is employed that uses combined 2D features for classifying images across COCO, Imagenet and SUN. The result of 2D CNN models is also compared with 1D LSTM and Bi-LSTM that utilizes raw fMRI BOLD signal for classification. It is seen that parallel CNN model outperforms other network models with an improvement of 7% for multiclass classification.

Electrical signals generated in the brain, known as Electroen-cephalographic (EEG) signals, are used in the study of the brain states spanning from normal wakefulness all the way to critical conditions such as seizure. This work aims to classify distinct EEG categories using EEG video representations with deep learning. EEG videos are utilized to obtain spatial, spectral and temporal features from EEG. The study utilizes Delaunay Triangulation interpolation to obtain EEG images from raw signals and further EEG videos are generated. EEG video features are given to CNN+LSTM network. Feature Pyramid Network (FPN) has also been adapted for EEG video classification. The results are presented here on two different classification scenarios that capture a range of cognitive activities: (1) EEG baselines (109 subjects) and (2) Mental Arithmetic vs Rest (36 subjects). EEG video-based analyses result in mean accuracies of 92.5% and 98.81% for the two different datasets with the improvement of 3.27% and 1.31% respectively over the state-of-the-art.

Electroencephalography (EEG) signals, recorded from different channels, are used to study human brain activity in the context of emotion recognition and seizure detection. Most of the existing emotion recognition methods have focused on EEG characteristics at an electrode level and not on connectivity patterns. Causal connectivity refers to the understanding of the causal relationship between the channels. In this work, we have developed an emotion recognition model using EEG-based causal connectivity patterns. Granger causality is used to find the causal relationship of the EEG signals from different channels. The quantification of causal configurations between the channels is carried out using Transfer Entropy. The obtained Transfer Entropy values are used as features for the classification of emotions. The performance of the proposed method is validated using a publicly available SEED-IV dataset. The proposed technique achieves an average subject-specific classification accuracy of 90 % (using 18 channel signals). The proposed method achieves an improvement of 1 % over state-of-the-art techniques based on correlation using 62 channel signals and an improvement of 17 % compared to methods that use only 18 channel signals.

This paper reports drone imagery-based precision agriculture application for coconut health management by detecting rhinoceros beetle infestation in coconut trees. Drone imagery is advantageous for its bird’s eye view of the farms helping in analysis of coconut crown from top. Locating and segmenting individual tree-crown is challenging, as every image contains up to 30 tree-crowns with complex backgrounds such as textured soil, shadows, companion planting. Dataset generated using 1,212 drone captured images containing 9727 individual coconut tree crowns. In this work, we are proposing enhancement to Invariant Risk Minimization (IRM) framework which is Signature Feature Marking (SFM) enhanced IRM for object classification. The proposed rhinoceros beetle infestation detection model is two stage process, (1) Applying existing IRM framework for crown detection and (2) SFM enhanced IRM for crown classification. IRM based crown detection model obtained 97.3% precision and 92% recall score and the SFM enhanced IRM classification model obtained accuracy of 85.03%. SFM learns relevant signature features from the images and IRM learns the causal correlation-based features. This demonstrates that drone imagery in precision agriculture using proposed approach can be effectively used to monitor the well-being of a large plantation.

A singular problem that mars the wide applicability of machine learning (ML) models is the lack of generalizability and interpretability. The ML community is increasingly working on bridging this gap. Prominent among them are methods that study causal significance of features, with techniques such as Average Causal Effect (ACE). In this paper, our objective is to utilize the causal analysis framework to measure the significance level of the features in binary classification task. Towards this, we propose a novel ACE-based metric called “Absolute area under ACE (A-ACE)” which computes the area of the absolute value of the ACE across different permissible levels of intervention. The performance of the proposed metric is illustrated on (i) ILSVRC (Imagenet) dataset and (ii) MNIST data set $(\sim 42000$ images) by considering pair-wise binary classification problem. Encouraging results have been observed on these two datasets. The computed metric values are found to be higher - peak performance of 10x higher than other for ILSVRC dataset and 50% higher than others for MNIST dataset - at precisely those locations that human intuition would mark as distinguishing regions. The method helps to capture the quantifiable metric which represents the distinction between the classes learnt by the model. This metric aids in visual explanation of the model’s prediction and thus, makes the model more trustworthy.

Glioma is a type of brain tumor that leaves the subject with very low survival rate. However the evidence showed that co-deletion of chromosome arms 1p/19q in low-grade glioma (LGG) revealed good response to therapy in LGG and is associated with longer survival rate. Therefore, predicting 1p/19q status is essential for effective treatment planning of LGG. There are several studies related with predicting deletion of chromosomal arms 1p/19q using multimodal medical images. Our study aims to classify 1p/19q co-deleted LGG status based on 3D volumetric probability distribution of glioma. Dataset are collected from TCIA public access. Study subjects included preoperative postcontrast-T1-W (T1C) and T2-W MRIs who had biopsy-proven 1p/19q status. A total of 159 grade-II and grade III LGG (57 non-deleted and 102 co-deleted) with 3 MRI slices in each subject were utilized in our study. The proposed method utilized statistical moments to obtain moment generating function (MGF) and characteristic function (CF). The optimal range of argument values were derived from MGF and CF for which the probability distribution of two classes were easily distinguishable. KS statistical test verified that the data of two classes have different distributions. Due to imbalanced dataset RUSboost classification was performed that yields better classification performance with MGF features compared to CF features. The average classification performance on the unseen test set with T1-W and T2-W MGF features were 93.22% (precision), 84.6% (recall), and 87.1% (accuracy) for grade-II and 90% (precision), 86% (recall), and 84% (accuracy) for grade-III.

The purpose of our study was to estimate the ability of structural MRI-based texture analysis (TA) to quantify the textural changes in corpus callosum (CC) caused by Progressive Supranuclear Palsy (PSP). Three distinct TA approaches: gray level co-occurrence matrix (GLCM), local binary pattern (LBP), oriented Gaussian derivative filter-bank (O-GDFB) were exploited and compared to present a novel study that can classify PSP ( n=16) by quantifying the textural changes in CC versus that in healthy controls (HC, n=16 using T1-W MRI. Different statistical texture features were extracted using GLCM and LBP, from the five distinct CC sub-regions: CC1 (genu), CC2 (anterior-mid), CC3 (mid-body), CC4 (posterior-mid) and CC5 (splenium). Unlike statistical based TA, a modified GDFB based TA was also performed to explore the differences in filter responses in the same CC sub-regions for different orientations and scales. Finally, texture features that showed statistically highest discrimibility were input to the SVM classifier in classifying PSP and HC. With leave-one-out cross validation, SVM classification results maximum accuracy of 88%, 94% and 84% at CC3 using GLCM, LBP and O-GDFB based TA respectively. The diagnostic potential of CC TA using structural MRI can be applied on routine radiology scan without needing any additional acquisition for quantitative analysis. Significant texture alteration found at CC3 may be of great importance in the diagnosis and understanding of this pathology.

Automatic eye screening for conditions like diabetic retinopathy critically hinges on detection and localization of Optic disk (OD). In this paper, we present a novel scale-embedded dictionary-based method that poses the problem of OD localization as that of classification, carried out in sparse representation framework. A dictionary is created with manually marked fixed-sized sub-images that contain OD at the center, for multiple scales. For a given test image, all subimages are sparsely represented as a linear combination of OD dictionary elements. A confidence measure indicating the likelihood of the presence of OD is obtained from these coefficients. Red channel and gray intensity images are processed independently, and their respective confidence measures are fused to form a confidence map. A blob detector is run on the confidence map, whose peak response is considered to be at the location of the OD. The proposed method is evaluated on publicly available databases such as DIARETDB0, DIARETDB1 and DRIVE. The OD was correctly localized in 253 out of 259 images, with an average computation time of 3.8 seconds/image and accuracy of 97.6%. Comparisons with two existing techniques are also discussed.

Magnetic resonance imaging of dynamic events such as cognitive tasks in the brain, requires high spatial and temporal resolution. In order to increase the resolution in both domains simultaneously, parallel imaging schemes have been in existence, where multiple receiver coils are used, each of which needs to acquire only a fraction of the total available signal. In our approach, we regularly undersample dersample the signal at each of the receiver coils and the resulting aliased coil images are combined (unaliased) using the neural network framework. Data acquisition follows a variable-density sampling scheme, where lower frequencies are densely sampled, and the remaining signal is sparsely sampled. The low resolution images obtained using the densely sampled low frequencies are used to train the neural network. Reconstruction of the image is carried out by feeding the high-resolution aliased images to the trained network. The proposed approach has been applied to phantom as well as real brain MRI data sets, and results have been compared with the standard existing parallel imaging techniques. The proposed approach is found to perform better than the standard existing techniques.