Research field of interests

Current research highlights

InVENTO: Intelligent Visual Encoder for Textual Output from Chest X-ray

Radiologist reports for X-Ray image-to-report generation:  we propose an Intelligent Visual Encoder for Textual Output (InVENTO) which has one lightweight transformer that takes the output from the frozen encoder and gives the most important text-based image embedding and bridges the modality gap between encoder and decoder using low resources. This is the two-step training approach: first, we train our intermediate Cross-Modal Query Fusion Layer (CMQFL) where we have learnable queries and in the second step, we fine-tune the decoder. This CMQFL is a lightweight transformer that significantly reduces the burden on the decoder in achieving alignment between vision and language.

AnoMed : Anomaly detection framework for multi-modal multi-organ medical images

AnoMed, a single-stage anchor-based mutual learning framework, utilizes Dense-Med—a detector with dual-attention guided Multi-level Multi-scale Feature Pyramidal Network. It tackle issues like inaccurate pseudo-labels and biased learning with a new Pseudo Label Strainer, incorporating a dynamic Load Balancing Scheme, and Adaptive Trainer System ensures stability through domain and distribution adaptation during semi-supervised learning. This certainly explores the potential of AI in computer aided diagnosis in current medical settings with greater speed and better accuracy.

Missed or delayed fracture detection from X-rays

We are currently working on the missed or delayed fracture detection to draw the attention of doctors to the affected area using deep learning. Also, we are working on the localization of the fracture to pinpoint the exact location of the fracture which can help the radiologist in accurate detection.

Automatic segmentation and annotation

Automatic segmentation and annotation of medical image plays a critical role. Automatic segmentation and annotation not only increase the efficiency of clinical workflow, but also prevent overburdening of radiologists. The objective of this work is to improve the accuracy and give probabilistic map for annotation.

Early Sepsis prediction

The main purpose of the project is to predict sepsis 6 hours earlier from patient’s physiological data. After getting the cleaned data we are using ML method to see their performance, but the accuracy is affected by class imbalance problem on which we are currently working.

Fine-grained visual classification

Fine-grained visual classification (FGVC) is challenging but more critical than traditional classification tasks. It requires distinguishing different subcategories with the inherently subtle intra-class object variations. Previous works focus on enhancing the feature representation ability using multiple granularities and discriminative regions based on the attention strategy or bounding boxes. However, these methods highly rely on deep neural networks which lack interpretability. Our target is to implement on medical images using some attention guided framework which can guide the network to extract discriminative regions in an interpretable way, a progressive training mechanism as the complexity of the problem, and transfer learning based on pretrained model on medical images.

Multimodality Breast Cancer Detection Using Artificial Intelligence

Many imaging modalities such as ultrasound, (X-ray) mammography, breast MRI are frequently used to examine the Breast Cancer severity. All imaging modalities have their pros and cons depending upon cost, image resolution, radiation uses etc. Further, early detection and disease severity prediction is a major challenge in medical practices.  Previous methods use machine learning based methods for classification of breast cancer into corresponding BI-RADS (Breast Imaging Reporting and Database system) scores. In the proposed work we intend to use deep learning-based methods for completely automated solution of breast cancer detection and disease severity classification. Further, based on the BI-RADS guidelines we suggest the patient for a biopsy of the diseased area. Once the biopsy images of the same patient are available, we’ll find the correlation between the histopathological images and corresponding radiological images. Finally, we’ll validate our system using some performance analysis and statistical tools.

Thoracic disease detection from a Chest X-ray image

The main purpose of this work is detecting thoracic disease from a Chest X-ray image and multiple disease from one image, if present.  We are currently working on visualizing the anomaly in parallel to disease detection which can further help the radiologist to better analyze and make concluding remarks.

Research Fund