Mustaffa Hussain

I'm a Senior Machine Learning Scientist at Onward Assist in hyderabad where I lead a small team that mostly works on developing PathAssist an AI Pathology Assistant.

Prior to that, I spent an year with Myways.ai where I worked on recommendation engine via knowledge graphs. I was at Machine Learning and Statistical Inference Lab (MLSI-LAB) from 2018-19 where I worked on Multi Task Learning Models and Extension.

I did my MS in Computer Science from South Asian University with specialisation in AI under Dr. Reshma Rastogi. I did my BS in Computer Science from Central University of Rajasthan.

Email  /  CV  /  Linkedin /  Github

I am interested in Multimodal vision language models, LLMs, Contrastive learning, Foundational models etc. Problems like distribution matching, sparse models for dense predictions, realtime machine vison, efficient 3D reconstruction excites me.

Having an IHC and H&E of a tissue helps in improving the cancer diagnostic procedure to a great extent. Having both types of stains comes with labor, space, monetary and accessablity etc issues. Merging the direct translations presents its own set of challenges at WSI level. We present an UNSB approach for unpaired image to image translation of IHC to H&E at WSI level for virtual staining.

Tags: Generative Image , Pytorch , Large Image Stitching

Functional tissue Units (FTUs) are cell population neighborhoods local to a particular organ performing its main function. We have developed a model to segment multi-organ FTUs across 5 organs namely: the kidney, large intestine, lung, prostate and spleen. We propose adding shifted auxiliary loss for optimal and robust training and present comprehensive results supporting our claim.

Tags: Image Segmentation , Auxiliary loss , Medical Image Analysis

Domain variability is a common bottle neck in developing generalisable algorithms for various medical applications. We propose to learn a target representative feature space through unpaired image to image translation (CycleGAN) and use translation to mitosis detection with candidate proposal and classification. This work presents a simple yet effective multi-step mitotic figure detection.

Tags: Object Detection , Domain Generalisation , RGB-HSV-LAB colour spaces

Multi-task learning performs better than the classical single task learning by learning from the training signals inherent in all the tasks. Inspired by multi-task twin support vector machine, we propose a novel robust multi-task least squares twin support vector machine for classification.

Tags: Twin SVM , Kernal Space , Hinge loss

I am blown by the pace at which AI is progressing. I started a community based tech blog space The CyPhy. It is a group of curious minds trying to understand and explain data better. Here are few articles in NLP, Vision and ML engineering domain written/moderated by me. Please visit The CyPhy for full list of articles.

Desiging a Knowledge Graph-based ChatBot

Dense Feature Detection in Satellite Imagery

Distribution matching measure in computer vision

Training Custom NER Model Using Flair

Handling Data Imbalance in Multi-label Classification

Deploying Deep Learning Model Using Docker, Tensorflow Serving, Nginx and Flask (Part1-3)

Compiled collection of resources that I find helpful . While some of the content may be outdated, I hope it is still useful.

Computer vision lectures

Foundations of Deep Learning