Sushovan Jena Email & Phone Number

Bengaluru, Ka, In

Bengaluru, Karnataka, India

Led the Data Science team for designing the end-end pipeline of AI based Cancer Detection solution (ASTRA) for real-time Whole Slide Image (WSI) Scanner (VisionX).Worked on Cervical cancer cytology high-resolution images, implementing a Multiple Instance Learning method (CHOWDER) for Weakly Supervised Learning where only WSI-level labels were available.For performance enhancement, introduced Self-Attention (Spatial and Channel Attention) on CHOWDER method and achieved 10% (approx.) improvement in Specificity and Sensitivity.To acheive better performance on Unseen domains, introduced Domain Invariance into the above model and also experimented with an existing paper RandStainNA for Stain Normalisation and Augmentation to achieve robustness to variation of stains in cytopathology.Started another research line for Moderately Supervised Learning on moderately annotated CPAP (Conventional PAP Smear) and LBC (Liquid based Cytology) images. Experimented with ResNet-18, ResNet-50 and ResNet-101 for tile-level classification and achieved 75% Sensitivity and 85% Specificity.

Now playing with Attentions spatially in High-resolution medical images.Worked on Unsupervised Anomaly Segmentation of Industrial Objects (MVTecAD dataset) using Knowledge Distillation and Quantization techniques.Implemented different Teacher-Student based CNN frameworks such as ResNet-18, WideResNet-50 for Unsupervised Anomaly Segmentation. Assessed their Multi-class Generalisation performance, Memory Footprints and Latency.Also assessed different Generative-based methods such as Variational Auto-encoders and GANs, and Memory-Module based techniques.Leveraged Knowledge Distillation, Post-Training Quantization and Quantization-aware Training across PyTorch and TensorRT frameworks to improve Latency on resource-constrained devices.Deployed models on NVIDIA Jetson Xavier NX using TensorRT and ONNX.Proposed methods to improve multi-class knowledge distillation using Spatial and Channel Attention with KL-Divergence and Cosine Distance losses.Tech invloved - Deep Learning, Computer Vision, Image Processing, Model Compression, Quantization, ONNX, Torchscript, Docker, PyTorch, TensorRT and obviously Python. Paper accepted and Presented in International Conference on Machine Vision 2023, International Conference on Pattern Recognition 2024.

Himachal Pradesh, India

TA for CS451-Computer Graphics and Game Design.TA for DS201-Data Handling and Visualisation

Mandi, Himachal Pradesh, India

Mandi, Himachal Pradesh, India

TA for Data science-III at IIT Mandi.

Chennai, Tamil Nadu, India

Developed an Automated Meter Reading model for Electricity meter images using Deep Learning models such as Faster-RCNN, Cascade Mask-RCNN using VGGNet and YOLOv3 and YOLOv6 on SAP AI Core. Leveraged Model Pruning for reducing Latency and Memory Footprint on SAP IoT apps for edge applications.Worked on SAP Cognitive Competency to implement AI and ML features to the end product for HR and Treasury Risk Management modules. It involved use-cases such as Workforce prediction and Forex predictions. It involved Feature Extraction, Dimensionality reduction, Polynomial Regression, Decision Trees.Developed SAP HANA Applications under some functional modules like Treasury and Risk Management, HR. The project is a part of "Quantum" - Building Wipro's own ERP on S/4 HANA, One of the largest project at Wipro.Collaborated with people of SAP India and built end-to-end objects from scratch with optimised SQL Queries. Also a part of Data Migration team and built codes through which large volumes of data can be transferred from SAP ECC to S/4 HANA Database following ETL standards.Tech invloved - Deep Learning, Computer Vision, Model Compression, OOPs, Advanced SQL, Database concepts, SAP OData (REST API).

Master Of Science - Ms (By Research), Computer Science

Thesis on - Anomaly Segmentation in Manufacturing and Industrial surfaces using Deep Learning, to be deployed on Edge devices. Core.

Bachelor Of Technology - Btech, Electrical And Electronics Engineering