Schizophrenia is a chronic cognitive disorder where clinical classification is challenging because of the lack of well-established, non-invasive diagnoses biomarkers. Hence, there is a need for objective systems that can classify schizophrenia using structural and functional Magnetic Resonance Imaging(MRI) components. We present experimentation to handle multimodal Functional Network Connectivity(FNC) and Source Based Morphometry(SBM) features in order to outperform machine learning benchmarks.

This concept revolving around multilabel classification uses chest radiographs to effectively diagnose Cardiomegaly, Atelectasis, Edema, Consolidation, and Plural effusion. An ensemble model with deep convolutional neural networks is used extracts features from various subsets of data.

The advent of neural networks has been a huge thrust to the automated healthcare sector and more so to an era of personalized medicine. In this light, the following work involving the automated detection of cancer from Whole Slide Images(WSI) which encompasses scanning through large swathes of benign regions to identify the areas of malignancy. The precise delineation of malignant tissue is crucial to the subsequent estimation of grading tumor aggressiveness. In this web application, we present a simple Convolutional Neural Network to automatically detect Invasive Ductal Carcinoma(IDC) from Breast HistoPathology Whole Slide Image(WSI) Patches. We experimented with six CNN models - where the number of parameters ranged from 50k to 20M - and we eventually choose a 300k parameter model that gave us the best performance(0.925 +- 0.005) for an optimal parameter count while ensuring stable convergence. The current focus is on fine-tuning these models to reduce risks and minimize error rates.

Generic enhancement over the Grad-CAM architecture provides improvised visual explanations of model predictions with better object localization extended to multiple object instances in a single image frame using weighted partial derivatives of previous convolutional layer feature maps with respect to specific score as generating visualizations of corresponding class labels for positive combinations.

A binary classifier is trained in order to effectively identify metastatic cancer in small image patches taken from larger digital pathology scans. The binary classifier relies on the idea of transfer learning implemented as Xception where existing convolutional neural network layers are reused to extract general feature information. The advantage of this algorithm specially with larger datasets like ImageNet is the ability to improvise the speed of the training process as we start with pre-trained weight matrix. In order to further enhance the performance, augmentation techniques are used. We obtained an accuracy rate of 94.91 when the model was trained for 90k images, where the validation set consisted of 10k images and the test set has 57k images with 10 epochs.

Gradient weighted class activation map interprets the intermediate results of a wide range of CNN architectures by using weights and distributions of any desired class flowing into the final convolutional layer to produce localisation maps of the significant regions of the image.

We're now having a new program catering to healthcare and development of AI specifically in these regions.