Harshit Pande

Click-through rate (CTR) prediction models are common in many online applications such as digital advertising and recommender systems. Field-Aware Factorization Machine (FFM) and Field-weighted Factorization Machine (FwFM) are state-of-the-art among the shallow models for CTR prediction. Recently, many deep learning-based models have also been proposed. Among deeper models, DeepFM, xDeepFM, AutoInt+, and FiBiNet are state-of-the-art models. The deeper models combine a core architectural… Show more

Click-through rate (CTR) prediction models are common in many online applications such as digital advertising and recommender systems. Field-Aware Factorization Machine (FFM) and Field-weighted Factorization Machine (FwFM) are state-of-the-art among the shallow models for CTR prediction. Recently, many deep learning-based models have also been proposed. Among deeper models, DeepFM, xDeepFM, AutoInt+, and FiBiNet are state-of-the-art models. The deeper models combine a core architectural component, which learns explicit feature interactions, with a deep neural network (DNN) component. We propose a novel shallow Field-Embedded Factorization Machine (FEFM) and its deep counterpart Deep Field-Embedded Factorization Machine (DeepFEFM). FEFM learns symmetric matrix embeddings for each field pair along with the usual single vector embeddings for each feature. FEFM has significantly lower model complexity than FFM and roughly the same complexity as FwFM. FEFM also has insightful mathematical properties about important fields and field interactions. DeepFEFM combines the FEFM interaction vectors learned by the FEFM component with a DNN and is thus able to learn higher order interactions. We conducted comprehensive experiments over a wide range of hyperparameters on two large publicly available real-world datasets. When comparing test AUC and log loss, the results show that FEFM and DeepFEFM outperform the existing state-of-the-art shallow and deep models for CTR prediction tasks. Show less

Deep learning-based grading of the fundus images of the retina is an active area of research. Various existing studies use different deep learning architectures on different datasets. Results of some of the studies could not be replicated in other studies. Thus a benchmarking study across multiple architectures spanning both classification and localization is needed. We present a comparative study of different state-of-the-art architectures trained on a proprietary dataset and tested on the… Show more

Deep learning-based grading of the fundus images of the retina is an active area of research. Various existing studies use different deep learning architectures on different datasets. Results of some of the studies could not be replicated in other studies. Thus a benchmarking study across multiple architectures spanning both classification and localization is needed. We present a comparative study of different state-of-the-art architectures trained on a proprietary dataset and tested on the publicly available Messidor-2 dataset. Although evidence is of utmost importance in AI-based medical diagnosis, most studies limit themselves to the classification performance and do not report the quantification of the performance of the abnormalities localization. To alleviate this, using class activation maps, we also report a comparison of localization scores for different architectures. For classification, we found that as the number of parameters increase, the models perform better, with NASNet yielding highest accuracy and average precision, recall, and F1-scores of around 95%. For localization, VGG19 outperformed all the models with a mean Intersection over Minimum of 0.45. We also found that there is a trade-off between classification performance and localization performance. As the models get deeper, their receptive field increases, causing them to perform well on classification but underperform on the localization of fine-grained abnormalities. Show less

Screening for the diagnosis of glaucoma through a fundus image can be determined by the optic cup to disc diameter ratio (CDR), which requires the segmentation of the cup and disc regions. In this paper, we propose two novel approaches, namely Parameter-Shared Branched Network (PSBN) and Weak Region of Interest Model-based segmentation (WRoIM) to identify disc and cup boundaries. Unlike the previous approaches, the proposed methods are trained endto-end through a single neural network… Show more

Screening for the diagnosis of glaucoma through a fundus image can be determined by the optic cup to disc diameter ratio (CDR), which requires the segmentation of the cup and disc regions. In this paper, we propose two novel approaches, namely Parameter-Shared Branched Network (PSBN) and Weak Region of Interest Model-based segmentation (WRoIM) to identify disc and cup boundaries. Unlike the previous approaches, the proposed methods are trained endto-end through a single neural network architecture and use dynamic cropping instead of manual or traditional computer vision-based cropping. We are able to achieve similar performance as that of state-of-the-art approaches with less number of network parameters. Our experiments include comparison with different best known methods on publicly available Drishti-GS1 and RIM-ONE v3 datasets. With 7.8 × 10 6 parameters our approach achieves a Dice score of 0.96/0.89 for disc/cup segmentation on Drishti-GS1 data whereas the existing state-of-the-art approach uses 19.8 × 10 6 parameters to achieve a dice score of 0.97/0.89. Show less

Autosomal recessive spastic ataxia of Charlevoix-Saguenay (ARSACS) is a neurodegenerative disease that is caused by mutations in the SACS gene. The product of this gene is a very large 520 kDa cytoplasmic protein, sacsin, with an ubiquitin-like domain (Ubl) at the N-terminus followed by three large sacsin internal repeats (SIRPTs) supradomains and C-terminal J and HEPN domains. The SIRPTs are predicted to contain Hsp90-like domains suggesting a potential chaperone activity. In this work, we… Show more

Autosomal recessive spastic ataxia of Charlevoix-Saguenay (ARSACS) is a neurodegenerative disease that is caused by mutations in the SACS gene. The product of this gene is a very large 520 kDa cytoplasmic protein, sacsin, with an ubiquitin-like domain (Ubl) at the N-terminus followed by three large sacsin internal repeats (SIRPTs) supradomains and C-terminal J and HEPN domains. The SIRPTs are predicted to contain Hsp90-like domains suggesting a potential chaperone activity. In this work, we report the structures of the Hsp90-like Sr1 domain of SIRPT1 and the N-terminal Ubl domain determined at 1.55 Å and 2.1 Å resolution, respectively. The Ubl domain crystallized as a swapped dimer that could be relevant in the context of full-length protein. The Sr1 domain displays the Bergerat protein fold with a characteristic nucleotide-binding pocket, though it binds nucleotides with very low affinity. The Sr1 structure reveals that ARSACS-causing missense mutations (R272H, R272C and T201K) disrupt protein folding most likely leading to sacsin degradation. This work lends a structural support to the view of sacsin as a molecular chaperone and provides a framework for future studies of this protein. Show less

We propose a novel unsupervised method for extracting objects from urine microscopy images and also applied U-net for extracting these objects. We fused these proposed methods with a known edge thresholding technique from an existing work on segmentation of urine microscopic images. Comparison between our proposed methods and the existing work showed that for certain object types the proposed unsupervised method with or without edge thresholding outperforms the other methods, while in other… Show more

We propose a novel unsupervised method for extracting objects from urine microscopy images and also applied U-net for extracting these objects. We fused these proposed methods with a known edge thresholding technique from an existing work on segmentation of urine microscopic images. Comparison between our proposed methods and the existing work showed that for certain object types the proposed unsupervised method with or without edge thresholding outperforms the other methods, while in other cases the U-net method with or without edge thresholding outperforms the other methods. Overall the proposed unsupervised method along with edge thresholding worked the best by extracting maximum number of objects and minimum number of artifacts. On a test dataset, the artifact to object ratio for the proposed unsupervised method was 0.71, which is significantly better than that of 1.26 for the existing work. The proposed unsupervised method along with edge thresholding extracted 3208 objects as compared to 1608 by the existing work. To the best of our knowledge this is the first application of Deep Learning for extraction of clinically significant objects in urine microscopy images. Show less

In this study, we evaluate the performance of Shrava, a cloud based Artificial Intelligence (AI) system for automated analysis of images captured from urine samples. Identification and morphological classification of objects in urine sediments by Shrava was compared with the results from Sysmex UF-1000i urine analyzer and manual microscopy. Thirty urine samples were analysed for the study, wherein, on an average, 50 different fields of views were captured at a magnification of 400x from slides… Show more

In this study, we evaluate the performance of Shrava, a cloud based Artificial Intelligence (AI) system for automated analysis of images captured from urine samples. Identification and morphological classification of objects in urine sediments by Shrava was compared with the results from Sysmex UF-1000i urine analyzer and manual microscopy. Thirty urine samples were analysed for the study, wherein, on an average, 50 different fields of views were captured at a magnification of 400x from slides prepared from the samples. Classification of objects from the captured images was verified by three qualified medical experts and sensitivity, specificity, and accuracy of the classification results were calculated. Classification performance of Shrava was evaluated for RBCs, WBCs, crystals, epithelial cells and organisms (yeast and bacteria). The specificity for classification was above 97% for RBCs and above 99% for all other objects, while sensitivity was above 99% for yeast and epithelial cells, above 97% for RBCs, WBCs, and bacteria, and above 87% for crystals. Overall, classification accuracy for all objects was 96.4%. We also evaluated the sensitivity of Shrava for the above mentioned objects vis-a-vis reports obtained through a combination of urine analyser and manual microscopy and it was found to be 96.19%. Shrava was found to be effective in identifying and classifying objects in urine sediments. It saves time by aiding pathologists as a screening solution and also accelerates the turnaround time, thereby, increasing the productivity of pathologists and the laboratory. Show less

We present a novel, unsupervised, and distance measure agnostic method for search space reduction in spell correction using neural character embeddings. The embeddings are learned by skip-gram word2vec training on sequences generated from dictionary words in a phonetic informationretentive manner. We report a very high performance in terms of both success rates and reduction of search space on the Birkbeck spelling error corpus.

Sacsin is a 520 kDa protein involved in an early onset neurodegenerative disease Autosomal Recessive Spastic Ataxia of Charlevoix-Saguenay (ARSACS) prevalent in the Charlevoix-Saguenay-Lac-Saint-Jean region of Quebec. Structural information about sacsin should help in clarifying the functional role of the protein and its involvement in neurodegeneration. The very C-terminus of the protein contains a Higher Eukaryotes and Prokaryotes Nucleotide-binding (HEPN) domain of unknown function. We… Show more

Sacsin is a 520 kDa protein involved in an early onset neurodegenerative disease Autosomal Recessive Spastic Ataxia of Charlevoix-Saguenay (ARSACS) prevalent in the Charlevoix-Saguenay-Lac-Saint-Jean region of Quebec. Structural information about sacsin should help in clarifying the functional role of the protein and its involvement in neurodegeneration. The very C-terminus of the protein contains a Higher Eukaryotes and Prokaryotes Nucleotide-binding (HEPN) domain of unknown function. We determined a high-resolution 1.9 Å crystal structure of the HEPN domain from human sacsin [1]. The structure is made of five helices in an antiparallel arrangement with a long loop between helices 4 and 5 containing several short structured segments. Importantly, the HEPN domain forms a stable dimer in the crystal with a large buried surface formed by helices 1 and 2 and the 4- 5 loop. Multi-angle light scattering and NMR self-diffusion experiments confirmed that the sacsin HEPN domain also forms a dimer in solution. The structure explains why the N4549D mutation causes disease in some ARSACS patients. The mutation leads to electrostatic repulsion near dimerization interface, destabilizing the dimer and resulting in insoluble protein. The HEPN structure contains a large positively charged cavity at the dimer interface that is optimized for binding negatively charged ligands. Isothermal titration calorimetry and NMR titrations showed that this surface binds nucleotides with low micromolar affinity, though the identity of physiological ligand is currently unclear. Recently, we crystallized and solved the structure of the N-terminal ubiquitin-like (UBL) domain from human sacsin at 2.1 Å resolution. Unexpectedly, the structure shows a swapped dimer with the swapping mediated by exchange of first -strands of each protomer. The study provides important steps towards better understanding of structure/function relationships of sacsin and its involvement in neurodegeneration. Show less

Autosomal recessive spastic ataxia of Charlevoix-Saguenay (ARSACS) is a progressive neurodegenerative disorder that first presents in early childhood, with prominent symptoms of spasticity in limbs, gait ataxia, slower motor development, muscle wasting and slurred speech. The disease is due to mutations in the SACS gene, which codes for sacsin, a large multidomain protein found in neurons. Loss of sacsin function is associated with progressive loss of cerebellar Purkinje cells and hyperfused… Show more

Autosomal recessive spastic ataxia of Charlevoix-Saguenay (ARSACS) is a progressive neurodegenerative disorder that first presents in early childhood, with prominent symptoms of spasticity in limbs, gait ataxia, slower motor development, muscle wasting and slurred speech. The disease is due to mutations in the SACS gene, which codes for sacsin, a large multidomain protein found in neurons. Loss of sacsin function is associated with progressive loss of cerebellar Purkinje cells and hyperfused mitochondrial network, most likely due to loss of the mitochondrial fission efficiency. Most of sacsin is structurally and functionally uncharacterized. Sacsin is believed to play a role as a chaperone for promoting the folding of ataxia-related proteins. Bioinformatics analysis showed that sacsin contains an integral ubiquitin-like domain (UBL) domain, which was subsequently shown to weakly interact with the proteasomal subunit C-8 in the coimmunoprecipitation studies. The research work described in this thesis includes the incorporation of selenomethionine in the UBL sequence, crystallization of the selenomethionine-labeled UBL domain to produce well-diffracting crystals, and determination of the structure of the UBL domain by using the anomalous scattering signal from selenium. The UBL structure obtained is unusual as it is a swapped dimer formed by the exchange of the N-terminal portions of two molecules. The existence of dimer was confirmed in solution by PFG-NMR self-diffusion experiments. The hydrophobic patch that is usually responsible for interaction of the UBL domain with other proteins is occluded in the swapped dimer, which suggests that the sacsin UBL domain does not bind the proteasome as a dimer. Show less